net/mlx5: separate the flow handle resource
[dpdk.git] / drivers / net / mlx5 / mlx5.c
1 /* SPDX-License-Identifier: BSD-3-Clause
2  * Copyright 2015 6WIND S.A.
3  * Copyright 2015 Mellanox Technologies, Ltd
4  */
5
6 #include <stddef.h>
7 #include <unistd.h>
8 #include <string.h>
9 #include <stdint.h>
10 #include <stdlib.h>
11 #include <errno.h>
12 #include <net/if.h>
13 #include <sys/mman.h>
14 #include <linux/rtnetlink.h>
15
16 /* Verbs header. */
17 /* ISO C doesn't support unnamed structs/unions, disabling -pedantic. */
18 #ifdef PEDANTIC
19 #pragma GCC diagnostic ignored "-Wpedantic"
20 #endif
21 #include <infiniband/verbs.h>
22 #ifdef PEDANTIC
23 #pragma GCC diagnostic error "-Wpedantic"
24 #endif
25
26 #include <rte_malloc.h>
27 #include <rte_ethdev_driver.h>
28 #include <rte_ethdev_pci.h>
29 #include <rte_pci.h>
30 #include <rte_bus_pci.h>
31 #include <rte_common.h>
32 #include <rte_kvargs.h>
33 #include <rte_rwlock.h>
34 #include <rte_spinlock.h>
35 #include <rte_string_fns.h>
36 #include <rte_alarm.h>
37
38 #include <mlx5_glue.h>
39 #include <mlx5_devx_cmds.h>
40 #include <mlx5_common.h>
41
42 #include "mlx5_defs.h"
43 #include "mlx5.h"
44 #include "mlx5_utils.h"
45 #include "mlx5_rxtx.h"
46 #include "mlx5_autoconf.h"
47 #include "mlx5_mr.h"
48 #include "mlx5_flow.h"
49 #include "rte_pmd_mlx5.h"
50
51 /* Device parameter to enable RX completion queue compression. */
52 #define MLX5_RXQ_CQE_COMP_EN "rxq_cqe_comp_en"
53
54 /* Device parameter to enable RX completion entry padding to 128B. */
55 #define MLX5_RXQ_CQE_PAD_EN "rxq_cqe_pad_en"
56
57 /* Device parameter to enable padding Rx packet to cacheline size. */
58 #define MLX5_RXQ_PKT_PAD_EN "rxq_pkt_pad_en"
59
60 /* Device parameter to enable Multi-Packet Rx queue. */
61 #define MLX5_RX_MPRQ_EN "mprq_en"
62
63 /* Device parameter to configure log 2 of the number of strides for MPRQ. */
64 #define MLX5_RX_MPRQ_LOG_STRIDE_NUM "mprq_log_stride_num"
65
66 /* Device parameter to limit the size of memcpy'd packet for MPRQ. */
67 #define MLX5_RX_MPRQ_MAX_MEMCPY_LEN "mprq_max_memcpy_len"
68
69 /* Device parameter to set the minimum number of Rx queues to enable MPRQ. */
70 #define MLX5_RXQS_MIN_MPRQ "rxqs_min_mprq"
71
72 /* Device parameter to configure inline send. Deprecated, ignored.*/
73 #define MLX5_TXQ_INLINE "txq_inline"
74
75 /* Device parameter to limit packet size to inline with ordinary SEND. */
76 #define MLX5_TXQ_INLINE_MAX "txq_inline_max"
77
78 /* Device parameter to configure minimal data size to inline. */
79 #define MLX5_TXQ_INLINE_MIN "txq_inline_min"
80
81 /* Device parameter to limit packet size to inline with Enhanced MPW. */
82 #define MLX5_TXQ_INLINE_MPW "txq_inline_mpw"
83
84 /*
85  * Device parameter to configure the number of TX queues threshold for
86  * enabling inline send.
87  */
88 #define MLX5_TXQS_MIN_INLINE "txqs_min_inline"
89
90 /*
91  * Device parameter to configure the number of TX queues threshold for
92  * enabling vectorized Tx, deprecated, ignored (no vectorized Tx routines).
93  */
94 #define MLX5_TXQS_MAX_VEC "txqs_max_vec"
95
96 /* Device parameter to enable multi-packet send WQEs. */
97 #define MLX5_TXQ_MPW_EN "txq_mpw_en"
98
99 /*
100  * Device parameter to force doorbell register mapping
101  * to non-cahed region eliminating the extra write memory barrier.
102  */
103 #define MLX5_TX_DB_NC "tx_db_nc"
104
105 /*
106  * Device parameter to include 2 dsegs in the title WQEBB.
107  * Deprecated, ignored.
108  */
109 #define MLX5_TXQ_MPW_HDR_DSEG_EN "txq_mpw_hdr_dseg_en"
110
111 /*
112  * Device parameter to limit the size of inlining packet.
113  * Deprecated, ignored.
114  */
115 #define MLX5_TXQ_MAX_INLINE_LEN "txq_max_inline_len"
116
117 /*
118  * Device parameter to enable hardware Tx vector.
119  * Deprecated, ignored (no vectorized Tx routines anymore).
120  */
121 #define MLX5_TX_VEC_EN "tx_vec_en"
122
123 /* Device parameter to enable hardware Rx vector. */
124 #define MLX5_RX_VEC_EN "rx_vec_en"
125
126 /* Allow L3 VXLAN flow creation. */
127 #define MLX5_L3_VXLAN_EN "l3_vxlan_en"
128
129 /* Activate DV E-Switch flow steering. */
130 #define MLX5_DV_ESW_EN "dv_esw_en"
131
132 /* Activate DV flow steering. */
133 #define MLX5_DV_FLOW_EN "dv_flow_en"
134
135 /* Enable extensive flow metadata support. */
136 #define MLX5_DV_XMETA_EN "dv_xmeta_en"
137
138 /* Activate Netlink support in VF mode. */
139 #define MLX5_VF_NL_EN "vf_nl_en"
140
141 /* Enable extending memsegs when creating a MR. */
142 #define MLX5_MR_EXT_MEMSEG_EN "mr_ext_memseg_en"
143
144 /* Select port representors to instantiate. */
145 #define MLX5_REPRESENTOR "representor"
146
147 /* Device parameter to configure the maximum number of dump files per queue. */
148 #define MLX5_MAX_DUMP_FILES_NUM "max_dump_files_num"
149
150 /* Configure timeout of LRO session (in microseconds). */
151 #define MLX5_LRO_TIMEOUT_USEC "lro_timeout_usec"
152
153 #ifndef HAVE_IBV_MLX5_MOD_MPW
154 #define MLX5DV_CONTEXT_FLAGS_MPW_ALLOWED (1 << 2)
155 #define MLX5DV_CONTEXT_FLAGS_ENHANCED_MPW (1 << 3)
156 #endif
157
158 #ifndef HAVE_IBV_MLX5_MOD_CQE_128B_COMP
159 #define MLX5DV_CONTEXT_FLAGS_CQE_128B_COMP (1 << 4)
160 #endif
161
162 static const char *MZ_MLX5_PMD_SHARED_DATA = "mlx5_pmd_shared_data";
163
164 /* Shared memory between primary and secondary processes. */
165 struct mlx5_shared_data *mlx5_shared_data;
166
167 /* Spinlock for mlx5_shared_data allocation. */
168 static rte_spinlock_t mlx5_shared_data_lock = RTE_SPINLOCK_INITIALIZER;
169
170 /* Process local data for secondary processes. */
171 static struct mlx5_local_data mlx5_local_data;
172
173 /** Driver-specific log messages type. */
174 int mlx5_logtype;
175
176 /** Data associated with devices to spawn. */
177 struct mlx5_dev_spawn_data {
178         uint32_t ifindex; /**< Network interface index. */
179         uint32_t max_port; /**< IB device maximal port index. */
180         uint32_t ibv_port; /**< IB device physical port index. */
181         int pf_bond; /**< bonding device PF index. < 0 - no bonding */
182         struct mlx5_switch_info info; /**< Switch information. */
183         struct ibv_device *ibv_dev; /**< Associated IB device. */
184         struct rte_eth_dev *eth_dev; /**< Associated Ethernet device. */
185         struct rte_pci_device *pci_dev; /**< Backend PCI device. */
186 };
187
188 static LIST_HEAD(, mlx5_ibv_shared) mlx5_ibv_list = LIST_HEAD_INITIALIZER();
189 static pthread_mutex_t mlx5_ibv_list_mutex = PTHREAD_MUTEX_INITIALIZER;
190
191 #define MLX5_FLOW_MIN_ID_POOL_SIZE 512
192 #define MLX5_ID_GENERATION_ARRAY_FACTOR 16
193
194 #define MLX5_FLOW_TABLE_HLIST_ARRAY_SIZE 4096
195 #define MLX5_TAGS_HLIST_ARRAY_SIZE 8192
196
197 /**
198  * Allocate ID pool structure.
199  *
200  * @param[in] max_id
201  *   The maximum id can be allocated from the pool.
202  *
203  * @return
204  *   Pointer to pool object, NULL value otherwise.
205  */
206 struct mlx5_flow_id_pool *
207 mlx5_flow_id_pool_alloc(uint32_t max_id)
208 {
209         struct mlx5_flow_id_pool *pool;
210         void *mem;
211
212         pool = rte_zmalloc("id pool allocation", sizeof(*pool),
213                            RTE_CACHE_LINE_SIZE);
214         if (!pool) {
215                 DRV_LOG(ERR, "can't allocate id pool");
216                 rte_errno  = ENOMEM;
217                 return NULL;
218         }
219         mem = rte_zmalloc("", MLX5_FLOW_MIN_ID_POOL_SIZE * sizeof(uint32_t),
220                           RTE_CACHE_LINE_SIZE);
221         if (!mem) {
222                 DRV_LOG(ERR, "can't allocate mem for id pool");
223                 rte_errno  = ENOMEM;
224                 goto error;
225         }
226         pool->free_arr = mem;
227         pool->curr = pool->free_arr;
228         pool->last = pool->free_arr + MLX5_FLOW_MIN_ID_POOL_SIZE;
229         pool->base_index = 0;
230         pool->max_id = max_id;
231         return pool;
232 error:
233         rte_free(pool);
234         return NULL;
235 }
236
237 /**
238  * Release ID pool structure.
239  *
240  * @param[in] pool
241  *   Pointer to flow id pool object to free.
242  */
243 void
244 mlx5_flow_id_pool_release(struct mlx5_flow_id_pool *pool)
245 {
246         rte_free(pool->free_arr);
247         rte_free(pool);
248 }
249
250 /**
251  * Generate ID.
252  *
253  * @param[in] pool
254  *   Pointer to flow id pool.
255  * @param[out] id
256  *   The generated ID.
257  *
258  * @return
259  *   0 on success, error value otherwise.
260  */
261 uint32_t
262 mlx5_flow_id_get(struct mlx5_flow_id_pool *pool, uint32_t *id)
263 {
264         if (pool->curr == pool->free_arr) {
265                 if (pool->base_index == pool->max_id) {
266                         rte_errno  = ENOMEM;
267                         DRV_LOG(ERR, "no free id");
268                         return -rte_errno;
269                 }
270                 *id = ++pool->base_index;
271                 return 0;
272         }
273         *id = *(--pool->curr);
274         return 0;
275 }
276
277 /**
278  * Release ID.
279  *
280  * @param[in] pool
281  *   Pointer to flow id pool.
282  * @param[out] id
283  *   The generated ID.
284  *
285  * @return
286  *   0 on success, error value otherwise.
287  */
288 uint32_t
289 mlx5_flow_id_release(struct mlx5_flow_id_pool *pool, uint32_t id)
290 {
291         uint32_t size;
292         uint32_t size2;
293         void *mem;
294
295         if (pool->curr == pool->last) {
296                 size = pool->curr - pool->free_arr;
297                 size2 = size * MLX5_ID_GENERATION_ARRAY_FACTOR;
298                 MLX5_ASSERT(size2 > size);
299                 mem = rte_malloc("", size2 * sizeof(uint32_t), 0);
300                 if (!mem) {
301                         DRV_LOG(ERR, "can't allocate mem for id pool");
302                         rte_errno  = ENOMEM;
303                         return -rte_errno;
304                 }
305                 memcpy(mem, pool->free_arr, size * sizeof(uint32_t));
306                 rte_free(pool->free_arr);
307                 pool->free_arr = mem;
308                 pool->curr = pool->free_arr + size;
309                 pool->last = pool->free_arr + size2;
310         }
311         *pool->curr = id;
312         pool->curr++;
313         return 0;
314 }
315
316 /**
317  * Initialize the counters management structure.
318  *
319  * @param[in] sh
320  *   Pointer to mlx5_ibv_shared object to free
321  */
322 static void
323 mlx5_flow_counters_mng_init(struct mlx5_ibv_shared *sh)
324 {
325         uint8_t i;
326
327         TAILQ_INIT(&sh->cmng.flow_counters);
328         for (i = 0; i < RTE_DIM(sh->cmng.ccont); ++i)
329                 TAILQ_INIT(&sh->cmng.ccont[i].pool_list);
330 }
331
332 /**
333  * Destroy all the resources allocated for a counter memory management.
334  *
335  * @param[in] mng
336  *   Pointer to the memory management structure.
337  */
338 static void
339 mlx5_flow_destroy_counter_stat_mem_mng(struct mlx5_counter_stats_mem_mng *mng)
340 {
341         uint8_t *mem = (uint8_t *)(uintptr_t)mng->raws[0].data;
342
343         LIST_REMOVE(mng, next);
344         claim_zero(mlx5_devx_cmd_destroy(mng->dm));
345         claim_zero(mlx5_glue->devx_umem_dereg(mng->umem));
346         rte_free(mem);
347 }
348
349 /**
350  * Close and release all the resources of the counters management.
351  *
352  * @param[in] sh
353  *   Pointer to mlx5_ibv_shared object to free.
354  */
355 static void
356 mlx5_flow_counters_mng_close(struct mlx5_ibv_shared *sh)
357 {
358         struct mlx5_counter_stats_mem_mng *mng;
359         uint8_t i;
360         int j;
361         int retries = 1024;
362
363         rte_errno = 0;
364         while (--retries) {
365                 rte_eal_alarm_cancel(mlx5_flow_query_alarm, sh);
366                 if (rte_errno != EINPROGRESS)
367                         break;
368                 rte_pause();
369         }
370         for (i = 0; i < RTE_DIM(sh->cmng.ccont); ++i) {
371                 struct mlx5_flow_counter_pool *pool;
372                 uint32_t batch = !!(i % 2);
373
374                 if (!sh->cmng.ccont[i].pools)
375                         continue;
376                 pool = TAILQ_FIRST(&sh->cmng.ccont[i].pool_list);
377                 while (pool) {
378                         if (batch) {
379                                 if (pool->min_dcs)
380                                         claim_zero
381                                         (mlx5_devx_cmd_destroy(pool->min_dcs));
382                         }
383                         for (j = 0; j < MLX5_COUNTERS_PER_POOL; ++j) {
384                                 if (pool->counters_raw[j].action)
385                                         claim_zero
386                                         (mlx5_glue->destroy_flow_action
387                                                (pool->counters_raw[j].action));
388                                 if (!batch && pool->counters_raw[j].dcs)
389                                         claim_zero(mlx5_devx_cmd_destroy
390                                                   (pool->counters_raw[j].dcs));
391                         }
392                         TAILQ_REMOVE(&sh->cmng.ccont[i].pool_list, pool,
393                                      next);
394                         rte_free(pool);
395                         pool = TAILQ_FIRST(&sh->cmng.ccont[i].pool_list);
396                 }
397                 rte_free(sh->cmng.ccont[i].pools);
398         }
399         mng = LIST_FIRST(&sh->cmng.mem_mngs);
400         while (mng) {
401                 mlx5_flow_destroy_counter_stat_mem_mng(mng);
402                 mng = LIST_FIRST(&sh->cmng.mem_mngs);
403         }
404         memset(&sh->cmng, 0, sizeof(sh->cmng));
405 }
406
407 /**
408  * Extract pdn of PD object using DV API.
409  *
410  * @param[in] pd
411  *   Pointer to the verbs PD object.
412  * @param[out] pdn
413  *   Pointer to the PD object number variable.
414  *
415  * @return
416  *   0 on success, error value otherwise.
417  */
418 #ifdef HAVE_IBV_FLOW_DV_SUPPORT
419 static int
420 mlx5_get_pdn(struct ibv_pd *pd __rte_unused, uint32_t *pdn __rte_unused)
421 {
422         struct mlx5dv_obj obj;
423         struct mlx5dv_pd pd_info;
424         int ret = 0;
425
426         obj.pd.in = pd;
427         obj.pd.out = &pd_info;
428         ret = mlx5_glue->dv_init_obj(&obj, MLX5DV_OBJ_PD);
429         if (ret) {
430                 DRV_LOG(DEBUG, "Fail to get PD object info");
431                 return ret;
432         }
433         *pdn = pd_info.pdn;
434         return 0;
435 }
436 #endif /* HAVE_IBV_FLOW_DV_SUPPORT */
437
438 static int
439 mlx5_config_doorbell_mapping_env(const struct mlx5_dev_config *config)
440 {
441         char *env;
442         int value;
443
444         MLX5_ASSERT(rte_eal_process_type() == RTE_PROC_PRIMARY);
445         /* Get environment variable to store. */
446         env = getenv(MLX5_SHUT_UP_BF);
447         value = env ? !!strcmp(env, "0") : MLX5_ARG_UNSET;
448         if (config->dbnc == MLX5_ARG_UNSET)
449                 setenv(MLX5_SHUT_UP_BF, MLX5_SHUT_UP_BF_DEFAULT, 1);
450         else
451                 setenv(MLX5_SHUT_UP_BF,
452                        config->dbnc == MLX5_TXDB_NCACHED ? "1" : "0", 1);
453         return value;
454 }
455
456 static void
457 mlx5_restore_doorbell_mapping_env(int value)
458 {
459         MLX5_ASSERT(rte_eal_process_type() == RTE_PROC_PRIMARY);
460         /* Restore the original environment variable state. */
461         if (value == MLX5_ARG_UNSET)
462                 unsetenv(MLX5_SHUT_UP_BF);
463         else
464                 setenv(MLX5_SHUT_UP_BF, value ? "1" : "0", 1);
465 }
466
467 /**
468  * Allocate shared IB device context. If there is multiport device the
469  * master and representors will share this context, if there is single
470  * port dedicated IB device, the context will be used by only given
471  * port due to unification.
472  *
473  * Routine first searches the context for the specified IB device name,
474  * if found the shared context assumed and reference counter is incremented.
475  * If no context found the new one is created and initialized with specified
476  * IB device context and parameters.
477  *
478  * @param[in] spawn
479  *   Pointer to the IB device attributes (name, port, etc).
480  * @param[in] config
481  *   Pointer to device configuration structure.
482  *
483  * @return
484  *   Pointer to mlx5_ibv_shared object on success,
485  *   otherwise NULL and rte_errno is set.
486  */
487 static struct mlx5_ibv_shared *
488 mlx5_alloc_shared_ibctx(const struct mlx5_dev_spawn_data *spawn,
489                         const struct mlx5_dev_config *config)
490 {
491         struct mlx5_ibv_shared *sh;
492         int dbmap_env;
493         int err = 0;
494         uint32_t i;
495 #ifdef HAVE_IBV_FLOW_DV_SUPPORT
496         struct mlx5_devx_tis_attr tis_attr = { 0 };
497 #endif
498
499         MLX5_ASSERT(spawn);
500         /* Secondary process should not create the shared context. */
501         MLX5_ASSERT(rte_eal_process_type() == RTE_PROC_PRIMARY);
502         pthread_mutex_lock(&mlx5_ibv_list_mutex);
503         /* Search for IB context by device name. */
504         LIST_FOREACH(sh, &mlx5_ibv_list, next) {
505                 if (!strcmp(sh->ibdev_name, spawn->ibv_dev->name)) {
506                         sh->refcnt++;
507                         goto exit;
508                 }
509         }
510         /* No device found, we have to create new shared context. */
511         MLX5_ASSERT(spawn->max_port);
512         sh = rte_zmalloc("ethdev shared ib context",
513                          sizeof(struct mlx5_ibv_shared) +
514                          spawn->max_port *
515                          sizeof(struct mlx5_ibv_shared_port),
516                          RTE_CACHE_LINE_SIZE);
517         if (!sh) {
518                 DRV_LOG(ERR, "shared context allocation failure");
519                 rte_errno  = ENOMEM;
520                 goto exit;
521         }
522         /*
523          * Configure environment variable "MLX5_BF_SHUT_UP"
524          * before the device creation. The rdma_core library
525          * checks the variable at device creation and
526          * stores the result internally.
527          */
528         dbmap_env = mlx5_config_doorbell_mapping_env(config);
529         /* Try to open IB device with DV first, then usual Verbs. */
530         errno = 0;
531         sh->ctx = mlx5_glue->dv_open_device(spawn->ibv_dev);
532         if (sh->ctx) {
533                 sh->devx = 1;
534                 DRV_LOG(DEBUG, "DevX is supported");
535                 /* The device is created, no need for environment. */
536                 mlx5_restore_doorbell_mapping_env(dbmap_env);
537         } else {
538                 /* The environment variable is still configured. */
539                 sh->ctx = mlx5_glue->open_device(spawn->ibv_dev);
540                 err = errno ? errno : ENODEV;
541                 /*
542                  * The environment variable is not needed anymore,
543                  * all device creation attempts are completed.
544                  */
545                 mlx5_restore_doorbell_mapping_env(dbmap_env);
546                 if (!sh->ctx)
547                         goto error;
548                 DRV_LOG(DEBUG, "DevX is NOT supported");
549         }
550         err = mlx5_glue->query_device_ex(sh->ctx, NULL, &sh->device_attr);
551         if (err) {
552                 DRV_LOG(DEBUG, "ibv_query_device_ex() failed");
553                 goto error;
554         }
555         sh->refcnt = 1;
556         sh->max_port = spawn->max_port;
557         strncpy(sh->ibdev_name, sh->ctx->device->name,
558                 sizeof(sh->ibdev_name));
559         strncpy(sh->ibdev_path, sh->ctx->device->ibdev_path,
560                 sizeof(sh->ibdev_path));
561         pthread_mutex_init(&sh->intr_mutex, NULL);
562         /*
563          * Setting port_id to max unallowed value means
564          * there is no interrupt subhandler installed for
565          * the given port index i.
566          */
567         for (i = 0; i < sh->max_port; i++) {
568                 sh->port[i].ih_port_id = RTE_MAX_ETHPORTS;
569                 sh->port[i].devx_ih_port_id = RTE_MAX_ETHPORTS;
570         }
571         sh->pd = mlx5_glue->alloc_pd(sh->ctx);
572         if (sh->pd == NULL) {
573                 DRV_LOG(ERR, "PD allocation failure");
574                 err = ENOMEM;
575                 goto error;
576         }
577 #ifdef HAVE_IBV_FLOW_DV_SUPPORT
578         if (sh->devx) {
579                 err = mlx5_get_pdn(sh->pd, &sh->pdn);
580                 if (err) {
581                         DRV_LOG(ERR, "Fail to extract pdn from PD");
582                         goto error;
583                 }
584                 sh->td = mlx5_devx_cmd_create_td(sh->ctx);
585                 if (!sh->td) {
586                         DRV_LOG(ERR, "TD allocation failure");
587                         err = ENOMEM;
588                         goto error;
589                 }
590                 tis_attr.transport_domain = sh->td->id;
591                 sh->tis = mlx5_devx_cmd_create_tis(sh->ctx, &tis_attr);
592                 if (!sh->tis) {
593                         DRV_LOG(ERR, "TIS allocation failure");
594                         err = ENOMEM;
595                         goto error;
596                 }
597         }
598         sh->flow_id_pool = mlx5_flow_id_pool_alloc(UINT32_MAX);
599         if (!sh->flow_id_pool) {
600                 DRV_LOG(ERR, "can't create flow id pool");
601                 err = ENOMEM;
602                 goto error;
603         }
604 #endif /* HAVE_IBV_FLOW_DV_SUPPORT */
605         /*
606          * Once the device is added to the list of memory event
607          * callback, its global MR cache table cannot be expanded
608          * on the fly because of deadlock. If it overflows, lookup
609          * should be done by searching MR list linearly, which is slow.
610          *
611          * At this point the device is not added to the memory
612          * event list yet, context is just being created.
613          */
614         err = mlx5_mr_btree_init(&sh->mr.cache,
615                                  MLX5_MR_BTREE_CACHE_N * 2,
616                                  spawn->pci_dev->device.numa_node);
617         if (err) {
618                 err = rte_errno;
619                 goto error;
620         }
621         mlx5_flow_counters_mng_init(sh);
622         /* Add device to memory callback list. */
623         rte_rwlock_write_lock(&mlx5_shared_data->mem_event_rwlock);
624         LIST_INSERT_HEAD(&mlx5_shared_data->mem_event_cb_list,
625                          sh, mem_event_cb);
626         rte_rwlock_write_unlock(&mlx5_shared_data->mem_event_rwlock);
627         /* Add context to the global device list. */
628         LIST_INSERT_HEAD(&mlx5_ibv_list, sh, next);
629 exit:
630         pthread_mutex_unlock(&mlx5_ibv_list_mutex);
631         return sh;
632 error:
633         pthread_mutex_unlock(&mlx5_ibv_list_mutex);
634         MLX5_ASSERT(sh);
635         if (sh->tis)
636                 claim_zero(mlx5_devx_cmd_destroy(sh->tis));
637         if (sh->td)
638                 claim_zero(mlx5_devx_cmd_destroy(sh->td));
639         if (sh->pd)
640                 claim_zero(mlx5_glue->dealloc_pd(sh->pd));
641         if (sh->ctx)
642                 claim_zero(mlx5_glue->close_device(sh->ctx));
643         if (sh->flow_id_pool)
644                 mlx5_flow_id_pool_release(sh->flow_id_pool);
645         rte_free(sh);
646         MLX5_ASSERT(err > 0);
647         rte_errno = err;
648         return NULL;
649 }
650
651 /**
652  * Free shared IB device context. Decrement counter and if zero free
653  * all allocated resources and close handles.
654  *
655  * @param[in] sh
656  *   Pointer to mlx5_ibv_shared object to free
657  */
658 static void
659 mlx5_free_shared_ibctx(struct mlx5_ibv_shared *sh)
660 {
661         pthread_mutex_lock(&mlx5_ibv_list_mutex);
662 #ifdef RTE_LIBRTE_MLX5_DEBUG
663         /* Check the object presence in the list. */
664         struct mlx5_ibv_shared *lctx;
665
666         LIST_FOREACH(lctx, &mlx5_ibv_list, next)
667                 if (lctx == sh)
668                         break;
669         MLX5_ASSERT(lctx);
670         if (lctx != sh) {
671                 DRV_LOG(ERR, "Freeing non-existing shared IB context");
672                 goto exit;
673         }
674 #endif
675         MLX5_ASSERT(sh);
676         MLX5_ASSERT(sh->refcnt);
677         /* Secondary process should not free the shared context. */
678         MLX5_ASSERT(rte_eal_process_type() == RTE_PROC_PRIMARY);
679         if (--sh->refcnt)
680                 goto exit;
681         /* Remove from memory callback device list. */
682         rte_rwlock_write_lock(&mlx5_shared_data->mem_event_rwlock);
683         LIST_REMOVE(sh, mem_event_cb);
684         rte_rwlock_write_unlock(&mlx5_shared_data->mem_event_rwlock);
685         /* Release created Memory Regions. */
686         mlx5_mr_release(sh);
687         /* Remove context from the global device list. */
688         LIST_REMOVE(sh, next);
689         /*
690          *  Ensure there is no async event handler installed.
691          *  Only primary process handles async device events.
692          **/
693         mlx5_flow_counters_mng_close(sh);
694         MLX5_ASSERT(!sh->intr_cnt);
695         if (sh->intr_cnt)
696                 mlx5_intr_callback_unregister
697                         (&sh->intr_handle, mlx5_dev_interrupt_handler, sh);
698 #ifdef HAVE_MLX5_DEVX_ASYNC_SUPPORT
699         if (sh->devx_intr_cnt) {
700                 if (sh->intr_handle_devx.fd)
701                         rte_intr_callback_unregister(&sh->intr_handle_devx,
702                                           mlx5_dev_interrupt_handler_devx, sh);
703                 if (sh->devx_comp)
704                         mlx5dv_devx_destroy_cmd_comp(sh->devx_comp);
705         }
706 #endif
707         pthread_mutex_destroy(&sh->intr_mutex);
708         if (sh->pd)
709                 claim_zero(mlx5_glue->dealloc_pd(sh->pd));
710         if (sh->tis)
711                 claim_zero(mlx5_devx_cmd_destroy(sh->tis));
712         if (sh->td)
713                 claim_zero(mlx5_devx_cmd_destroy(sh->td));
714         if (sh->ctx)
715                 claim_zero(mlx5_glue->close_device(sh->ctx));
716         if (sh->flow_id_pool)
717                 mlx5_flow_id_pool_release(sh->flow_id_pool);
718         rte_free(sh);
719 exit:
720         pthread_mutex_unlock(&mlx5_ibv_list_mutex);
721 }
722
723 /**
724  * Destroy table hash list and all the root entries per domain.
725  *
726  * @param[in] priv
727  *   Pointer to the private device data structure.
728  */
729 static void
730 mlx5_free_table_hash_list(struct mlx5_priv *priv)
731 {
732         struct mlx5_ibv_shared *sh = priv->sh;
733         struct mlx5_flow_tbl_data_entry *tbl_data;
734         union mlx5_flow_tbl_key table_key = {
735                 {
736                         .table_id = 0,
737                         .reserved = 0,
738                         .domain = 0,
739                         .direction = 0,
740                 }
741         };
742         struct mlx5_hlist_entry *pos;
743
744         if (!sh->flow_tbls)
745                 return;
746         pos = mlx5_hlist_lookup(sh->flow_tbls, table_key.v64);
747         if (pos) {
748                 tbl_data = container_of(pos, struct mlx5_flow_tbl_data_entry,
749                                         entry);
750                 MLX5_ASSERT(tbl_data);
751                 mlx5_hlist_remove(sh->flow_tbls, pos);
752                 rte_free(tbl_data);
753         }
754         table_key.direction = 1;
755         pos = mlx5_hlist_lookup(sh->flow_tbls, table_key.v64);
756         if (pos) {
757                 tbl_data = container_of(pos, struct mlx5_flow_tbl_data_entry,
758                                         entry);
759                 MLX5_ASSERT(tbl_data);
760                 mlx5_hlist_remove(sh->flow_tbls, pos);
761                 rte_free(tbl_data);
762         }
763         table_key.direction = 0;
764         table_key.domain = 1;
765         pos = mlx5_hlist_lookup(sh->flow_tbls, table_key.v64);
766         if (pos) {
767                 tbl_data = container_of(pos, struct mlx5_flow_tbl_data_entry,
768                                         entry);
769                 MLX5_ASSERT(tbl_data);
770                 mlx5_hlist_remove(sh->flow_tbls, pos);
771                 rte_free(tbl_data);
772         }
773         mlx5_hlist_destroy(sh->flow_tbls, NULL, NULL);
774 }
775
776 /**
777  * Initialize flow table hash list and create the root tables entry
778  * for each domain.
779  *
780  * @param[in] priv
781  *   Pointer to the private device data structure.
782  *
783  * @return
784  *   Zero on success, positive error code otherwise.
785  */
786 static int
787 mlx5_alloc_table_hash_list(struct mlx5_priv *priv)
788 {
789         struct mlx5_ibv_shared *sh = priv->sh;
790         char s[MLX5_HLIST_NAMESIZE];
791         int err = 0;
792
793         MLX5_ASSERT(sh);
794         snprintf(s, sizeof(s), "%s_flow_table", priv->sh->ibdev_name);
795         sh->flow_tbls = mlx5_hlist_create(s, MLX5_FLOW_TABLE_HLIST_ARRAY_SIZE);
796         if (!sh->flow_tbls) {
797                 DRV_LOG(ERR, "flow tables with hash creation failed.\n");
798                 err = ENOMEM;
799                 return err;
800         }
801 #ifndef HAVE_MLX5DV_DR
802         /*
803          * In case we have not DR support, the zero tables should be created
804          * because DV expect to see them even if they cannot be created by
805          * RDMA-CORE.
806          */
807         union mlx5_flow_tbl_key table_key = {
808                 {
809                         .table_id = 0,
810                         .reserved = 0,
811                         .domain = 0,
812                         .direction = 0,
813                 }
814         };
815         struct mlx5_flow_tbl_data_entry *tbl_data = rte_zmalloc(NULL,
816                                                           sizeof(*tbl_data), 0);
817
818         if (!tbl_data) {
819                 err = ENOMEM;
820                 goto error;
821         }
822         tbl_data->entry.key = table_key.v64;
823         err = mlx5_hlist_insert(sh->flow_tbls, &tbl_data->entry);
824         if (err)
825                 goto error;
826         rte_atomic32_init(&tbl_data->tbl.refcnt);
827         rte_atomic32_inc(&tbl_data->tbl.refcnt);
828         table_key.direction = 1;
829         tbl_data = rte_zmalloc(NULL, sizeof(*tbl_data), 0);
830         if (!tbl_data) {
831                 err = ENOMEM;
832                 goto error;
833         }
834         tbl_data->entry.key = table_key.v64;
835         err = mlx5_hlist_insert(sh->flow_tbls, &tbl_data->entry);
836         if (err)
837                 goto error;
838         rte_atomic32_init(&tbl_data->tbl.refcnt);
839         rte_atomic32_inc(&tbl_data->tbl.refcnt);
840         table_key.direction = 0;
841         table_key.domain = 1;
842         tbl_data = rte_zmalloc(NULL, sizeof(*tbl_data), 0);
843         if (!tbl_data) {
844                 err = ENOMEM;
845                 goto error;
846         }
847         tbl_data->entry.key = table_key.v64;
848         err = mlx5_hlist_insert(sh->flow_tbls, &tbl_data->entry);
849         if (err)
850                 goto error;
851         rte_atomic32_init(&tbl_data->tbl.refcnt);
852         rte_atomic32_inc(&tbl_data->tbl.refcnt);
853         return err;
854 error:
855         mlx5_free_table_hash_list(priv);
856 #endif /* HAVE_MLX5DV_DR */
857         return err;
858 }
859
860 /**
861  * Initialize DR related data within private structure.
862  * Routine checks the reference counter and does actual
863  * resources creation/initialization only if counter is zero.
864  *
865  * @param[in] priv
866  *   Pointer to the private device data structure.
867  *
868  * @return
869  *   Zero on success, positive error code otherwise.
870  */
871 static int
872 mlx5_alloc_shared_dr(struct mlx5_priv *priv)
873 {
874         struct mlx5_ibv_shared *sh = priv->sh;
875         char s[MLX5_HLIST_NAMESIZE];
876         int err = 0;
877
878         if (!sh->flow_tbls)
879                 err = mlx5_alloc_table_hash_list(priv);
880         else
881                 DRV_LOG(DEBUG, "sh->flow_tbls[%p] already created, reuse\n",
882                         (void *)sh->flow_tbls);
883         if (err)
884                 return err;
885         /* Create tags hash list table. */
886         snprintf(s, sizeof(s), "%s_tags", sh->ibdev_name);
887         sh->tag_table = mlx5_hlist_create(s, MLX5_TAGS_HLIST_ARRAY_SIZE);
888         if (!sh->tag_table) {
889                 DRV_LOG(ERR, "tags with hash creation failed.\n");
890                 err = ENOMEM;
891                 goto error;
892         }
893 #ifdef HAVE_MLX5DV_DR
894         void *domain;
895
896         if (sh->dv_refcnt) {
897                 /* Shared DV/DR structures is already initialized. */
898                 sh->dv_refcnt++;
899                 priv->dr_shared = 1;
900                 return 0;
901         }
902         /* Reference counter is zero, we should initialize structures. */
903         domain = mlx5_glue->dr_create_domain(sh->ctx,
904                                              MLX5DV_DR_DOMAIN_TYPE_NIC_RX);
905         if (!domain) {
906                 DRV_LOG(ERR, "ingress mlx5dv_dr_create_domain failed");
907                 err = errno;
908                 goto error;
909         }
910         sh->rx_domain = domain;
911         domain = mlx5_glue->dr_create_domain(sh->ctx,
912                                              MLX5DV_DR_DOMAIN_TYPE_NIC_TX);
913         if (!domain) {
914                 DRV_LOG(ERR, "egress mlx5dv_dr_create_domain failed");
915                 err = errno;
916                 goto error;
917         }
918         pthread_mutex_init(&sh->dv_mutex, NULL);
919         sh->tx_domain = domain;
920 #ifdef HAVE_MLX5DV_DR_ESWITCH
921         if (priv->config.dv_esw_en) {
922                 domain  = mlx5_glue->dr_create_domain
923                         (sh->ctx, MLX5DV_DR_DOMAIN_TYPE_FDB);
924                 if (!domain) {
925                         DRV_LOG(ERR, "FDB mlx5dv_dr_create_domain failed");
926                         err = errno;
927                         goto error;
928                 }
929                 sh->fdb_domain = domain;
930                 sh->esw_drop_action = mlx5_glue->dr_create_flow_action_drop();
931         }
932 #endif
933         sh->pop_vlan_action = mlx5_glue->dr_create_flow_action_pop_vlan();
934 #endif /* HAVE_MLX5DV_DR */
935         sh->dv_refcnt++;
936         priv->dr_shared = 1;
937         return 0;
938 error:
939         /* Rollback the created objects. */
940         if (sh->rx_domain) {
941                 mlx5_glue->dr_destroy_domain(sh->rx_domain);
942                 sh->rx_domain = NULL;
943         }
944         if (sh->tx_domain) {
945                 mlx5_glue->dr_destroy_domain(sh->tx_domain);
946                 sh->tx_domain = NULL;
947         }
948         if (sh->fdb_domain) {
949                 mlx5_glue->dr_destroy_domain(sh->fdb_domain);
950                 sh->fdb_domain = NULL;
951         }
952         if (sh->esw_drop_action) {
953                 mlx5_glue->destroy_flow_action(sh->esw_drop_action);
954                 sh->esw_drop_action = NULL;
955         }
956         if (sh->pop_vlan_action) {
957                 mlx5_glue->destroy_flow_action(sh->pop_vlan_action);
958                 sh->pop_vlan_action = NULL;
959         }
960         if (sh->tag_table) {
961                 /* tags should be destroyed with flow before. */
962                 mlx5_hlist_destroy(sh->tag_table, NULL, NULL);
963                 sh->tag_table = NULL;
964         }
965         mlx5_free_table_hash_list(priv);
966         return err;
967 }
968
969 /**
970  * Destroy DR related data within private structure.
971  *
972  * @param[in] priv
973  *   Pointer to the private device data structure.
974  */
975 static void
976 mlx5_free_shared_dr(struct mlx5_priv *priv)
977 {
978         struct mlx5_ibv_shared *sh;
979
980         if (!priv->dr_shared)
981                 return;
982         priv->dr_shared = 0;
983         sh = priv->sh;
984         MLX5_ASSERT(sh);
985 #ifdef HAVE_MLX5DV_DR
986         MLX5_ASSERT(sh->dv_refcnt);
987         if (sh->dv_refcnt && --sh->dv_refcnt)
988                 return;
989         if (sh->rx_domain) {
990                 mlx5_glue->dr_destroy_domain(sh->rx_domain);
991                 sh->rx_domain = NULL;
992         }
993         if (sh->tx_domain) {
994                 mlx5_glue->dr_destroy_domain(sh->tx_domain);
995                 sh->tx_domain = NULL;
996         }
997 #ifdef HAVE_MLX5DV_DR_ESWITCH
998         if (sh->fdb_domain) {
999                 mlx5_glue->dr_destroy_domain(sh->fdb_domain);
1000                 sh->fdb_domain = NULL;
1001         }
1002         if (sh->esw_drop_action) {
1003                 mlx5_glue->destroy_flow_action(sh->esw_drop_action);
1004                 sh->esw_drop_action = NULL;
1005         }
1006 #endif
1007         if (sh->pop_vlan_action) {
1008                 mlx5_glue->destroy_flow_action(sh->pop_vlan_action);
1009                 sh->pop_vlan_action = NULL;
1010         }
1011         pthread_mutex_destroy(&sh->dv_mutex);
1012 #endif /* HAVE_MLX5DV_DR */
1013         if (sh->tag_table) {
1014                 /* tags should be destroyed with flow before. */
1015                 mlx5_hlist_destroy(sh->tag_table, NULL, NULL);
1016                 sh->tag_table = NULL;
1017         }
1018         mlx5_free_table_hash_list(priv);
1019 }
1020
1021 /**
1022  * Initialize shared data between primary and secondary process.
1023  *
1024  * A memzone is reserved by primary process and secondary processes attach to
1025  * the memzone.
1026  *
1027  * @return
1028  *   0 on success, a negative errno value otherwise and rte_errno is set.
1029  */
1030 static int
1031 mlx5_init_shared_data(void)
1032 {
1033         const struct rte_memzone *mz;
1034         int ret = 0;
1035
1036         rte_spinlock_lock(&mlx5_shared_data_lock);
1037         if (mlx5_shared_data == NULL) {
1038                 if (rte_eal_process_type() == RTE_PROC_PRIMARY) {
1039                         /* Allocate shared memory. */
1040                         mz = rte_memzone_reserve(MZ_MLX5_PMD_SHARED_DATA,
1041                                                  sizeof(*mlx5_shared_data),
1042                                                  SOCKET_ID_ANY, 0);
1043                         if (mz == NULL) {
1044                                 DRV_LOG(ERR,
1045                                         "Cannot allocate mlx5 shared data");
1046                                 ret = -rte_errno;
1047                                 goto error;
1048                         }
1049                         mlx5_shared_data = mz->addr;
1050                         memset(mlx5_shared_data, 0, sizeof(*mlx5_shared_data));
1051                         rte_spinlock_init(&mlx5_shared_data->lock);
1052                 } else {
1053                         /* Lookup allocated shared memory. */
1054                         mz = rte_memzone_lookup(MZ_MLX5_PMD_SHARED_DATA);
1055                         if (mz == NULL) {
1056                                 DRV_LOG(ERR,
1057                                         "Cannot attach mlx5 shared data");
1058                                 ret = -rte_errno;
1059                                 goto error;
1060                         }
1061                         mlx5_shared_data = mz->addr;
1062                         memset(&mlx5_local_data, 0, sizeof(mlx5_local_data));
1063                 }
1064         }
1065 error:
1066         rte_spinlock_unlock(&mlx5_shared_data_lock);
1067         return ret;
1068 }
1069
1070 /**
1071  * Retrieve integer value from environment variable.
1072  *
1073  * @param[in] name
1074  *   Environment variable name.
1075  *
1076  * @return
1077  *   Integer value, 0 if the variable is not set.
1078  */
1079 int
1080 mlx5_getenv_int(const char *name)
1081 {
1082         const char *val = getenv(name);
1083
1084         if (val == NULL)
1085                 return 0;
1086         return atoi(val);
1087 }
1088
1089 /**
1090  * Verbs callback to allocate a memory. This function should allocate the space
1091  * according to the size provided residing inside a huge page.
1092  * Please note that all allocation must respect the alignment from libmlx5
1093  * (i.e. currently sysconf(_SC_PAGESIZE)).
1094  *
1095  * @param[in] size
1096  *   The size in bytes of the memory to allocate.
1097  * @param[in] data
1098  *   A pointer to the callback data.
1099  *
1100  * @return
1101  *   Allocated buffer, NULL otherwise and rte_errno is set.
1102  */
1103 static void *
1104 mlx5_alloc_verbs_buf(size_t size, void *data)
1105 {
1106         struct mlx5_priv *priv = data;
1107         void *ret;
1108         size_t alignment = sysconf(_SC_PAGESIZE);
1109         unsigned int socket = SOCKET_ID_ANY;
1110
1111         if (priv->verbs_alloc_ctx.type == MLX5_VERBS_ALLOC_TYPE_TX_QUEUE) {
1112                 const struct mlx5_txq_ctrl *ctrl = priv->verbs_alloc_ctx.obj;
1113
1114                 socket = ctrl->socket;
1115         } else if (priv->verbs_alloc_ctx.type ==
1116                    MLX5_VERBS_ALLOC_TYPE_RX_QUEUE) {
1117                 const struct mlx5_rxq_ctrl *ctrl = priv->verbs_alloc_ctx.obj;
1118
1119                 socket = ctrl->socket;
1120         }
1121         MLX5_ASSERT(data != NULL);
1122         ret = rte_malloc_socket(__func__, size, alignment, socket);
1123         if (!ret && size)
1124                 rte_errno = ENOMEM;
1125         return ret;
1126 }
1127
1128 /**
1129  * Verbs callback to free a memory.
1130  *
1131  * @param[in] ptr
1132  *   A pointer to the memory to free.
1133  * @param[in] data
1134  *   A pointer to the callback data.
1135  */
1136 static void
1137 mlx5_free_verbs_buf(void *ptr, void *data __rte_unused)
1138 {
1139         MLX5_ASSERT(data != NULL);
1140         rte_free(ptr);
1141 }
1142
1143 /**
1144  * DPDK callback to add udp tunnel port
1145  *
1146  * @param[in] dev
1147  *   A pointer to eth_dev
1148  * @param[in] udp_tunnel
1149  *   A pointer to udp tunnel
1150  *
1151  * @return
1152  *   0 on valid udp ports and tunnels, -ENOTSUP otherwise.
1153  */
1154 int
1155 mlx5_udp_tunnel_port_add(struct rte_eth_dev *dev __rte_unused,
1156                          struct rte_eth_udp_tunnel *udp_tunnel)
1157 {
1158         MLX5_ASSERT(udp_tunnel != NULL);
1159         if (udp_tunnel->prot_type == RTE_TUNNEL_TYPE_VXLAN &&
1160             udp_tunnel->udp_port == 4789)
1161                 return 0;
1162         if (udp_tunnel->prot_type == RTE_TUNNEL_TYPE_VXLAN_GPE &&
1163             udp_tunnel->udp_port == 4790)
1164                 return 0;
1165         return -ENOTSUP;
1166 }
1167
1168 /**
1169  * Initialize process private data structure.
1170  *
1171  * @param dev
1172  *   Pointer to Ethernet device structure.
1173  *
1174  * @return
1175  *   0 on success, a negative errno value otherwise and rte_errno is set.
1176  */
1177 int
1178 mlx5_proc_priv_init(struct rte_eth_dev *dev)
1179 {
1180         struct mlx5_priv *priv = dev->data->dev_private;
1181         struct mlx5_proc_priv *ppriv;
1182         size_t ppriv_size;
1183
1184         /*
1185          * UAR register table follows the process private structure. BlueFlame
1186          * registers for Tx queues are stored in the table.
1187          */
1188         ppriv_size =
1189                 sizeof(struct mlx5_proc_priv) + priv->txqs_n * sizeof(void *);
1190         ppriv = rte_malloc_socket("mlx5_proc_priv", ppriv_size,
1191                                   RTE_CACHE_LINE_SIZE, dev->device->numa_node);
1192         if (!ppriv) {
1193                 rte_errno = ENOMEM;
1194                 return -rte_errno;
1195         }
1196         ppriv->uar_table_sz = ppriv_size;
1197         dev->process_private = ppriv;
1198         return 0;
1199 }
1200
1201 /**
1202  * Un-initialize process private data structure.
1203  *
1204  * @param dev
1205  *   Pointer to Ethernet device structure.
1206  */
1207 static void
1208 mlx5_proc_priv_uninit(struct rte_eth_dev *dev)
1209 {
1210         if (!dev->process_private)
1211                 return;
1212         rte_free(dev->process_private);
1213         dev->process_private = NULL;
1214 }
1215
1216 /**
1217  * DPDK callback to close the device.
1218  *
1219  * Destroy all queues and objects, free memory.
1220  *
1221  * @param dev
1222  *   Pointer to Ethernet device structure.
1223  */
1224 static void
1225 mlx5_dev_close(struct rte_eth_dev *dev)
1226 {
1227         struct mlx5_priv *priv = dev->data->dev_private;
1228         unsigned int i;
1229         int ret;
1230
1231         DRV_LOG(DEBUG, "port %u closing device \"%s\"",
1232                 dev->data->port_id,
1233                 ((priv->sh->ctx != NULL) ? priv->sh->ctx->device->name : ""));
1234         /* In case mlx5_dev_stop() has not been called. */
1235         mlx5_dev_interrupt_handler_uninstall(dev);
1236         mlx5_dev_interrupt_handler_devx_uninstall(dev);
1237         /*
1238          * If default mreg copy action is removed at the stop stage,
1239          * the search will return none and nothing will be done anymore.
1240          */
1241         mlx5_flow_stop_default(dev);
1242         mlx5_traffic_disable(dev);
1243         /*
1244          * If all the flows are already flushed in the device stop stage,
1245          * then this will return directly without any action.
1246          */
1247         mlx5_flow_list_flush(dev, &priv->flows, true);
1248         mlx5_flow_meter_flush(dev, NULL);
1249         /* Free the intermediate buffers for flow creation. */
1250         mlx5_flow_free_intermediate(dev);
1251         /* Prevent crashes when queues are still in use. */
1252         dev->rx_pkt_burst = removed_rx_burst;
1253         dev->tx_pkt_burst = removed_tx_burst;
1254         rte_wmb();
1255         /* Disable datapath on secondary process. */
1256         mlx5_mp_req_stop_rxtx(dev);
1257         if (priv->rxqs != NULL) {
1258                 /* XXX race condition if mlx5_rx_burst() is still running. */
1259                 usleep(1000);
1260                 for (i = 0; (i != priv->rxqs_n); ++i)
1261                         mlx5_rxq_release(dev, i);
1262                 priv->rxqs_n = 0;
1263                 priv->rxqs = NULL;
1264         }
1265         if (priv->txqs != NULL) {
1266                 /* XXX race condition if mlx5_tx_burst() is still running. */
1267                 usleep(1000);
1268                 for (i = 0; (i != priv->txqs_n); ++i)
1269                         mlx5_txq_release(dev, i);
1270                 priv->txqs_n = 0;
1271                 priv->txqs = NULL;
1272         }
1273         mlx5_proc_priv_uninit(dev);
1274         if (priv->mreg_cp_tbl)
1275                 mlx5_hlist_destroy(priv->mreg_cp_tbl, NULL, NULL);
1276         mlx5_mprq_free_mp(dev);
1277         mlx5_free_shared_dr(priv);
1278         if (priv->rss_conf.rss_key != NULL)
1279                 rte_free(priv->rss_conf.rss_key);
1280         if (priv->reta_idx != NULL)
1281                 rte_free(priv->reta_idx);
1282         if (priv->config.vf)
1283                 mlx5_nl_mac_addr_flush(priv->nl_socket_route, mlx5_ifindex(dev),
1284                                        dev->data->mac_addrs,
1285                                        MLX5_MAX_MAC_ADDRESSES, priv->mac_own);
1286         if (priv->nl_socket_route >= 0)
1287                 close(priv->nl_socket_route);
1288         if (priv->nl_socket_rdma >= 0)
1289                 close(priv->nl_socket_rdma);
1290         if (priv->vmwa_context)
1291                 mlx5_vlan_vmwa_exit(priv->vmwa_context);
1292         if (priv->sh) {
1293                 /*
1294                  * Free the shared context in last turn, because the cleanup
1295                  * routines above may use some shared fields, like
1296                  * mlx5_nl_mac_addr_flush() uses ibdev_path for retrieveing
1297                  * ifindex if Netlink fails.
1298                  */
1299                 mlx5_free_shared_ibctx(priv->sh);
1300                 priv->sh = NULL;
1301         }
1302         ret = mlx5_hrxq_verify(dev);
1303         if (ret)
1304                 DRV_LOG(WARNING, "port %u some hash Rx queue still remain",
1305                         dev->data->port_id);
1306         ret = mlx5_ind_table_obj_verify(dev);
1307         if (ret)
1308                 DRV_LOG(WARNING, "port %u some indirection table still remain",
1309                         dev->data->port_id);
1310         ret = mlx5_rxq_obj_verify(dev);
1311         if (ret)
1312                 DRV_LOG(WARNING, "port %u some Rx queue objects still remain",
1313                         dev->data->port_id);
1314         ret = mlx5_rxq_verify(dev);
1315         if (ret)
1316                 DRV_LOG(WARNING, "port %u some Rx queues still remain",
1317                         dev->data->port_id);
1318         ret = mlx5_txq_obj_verify(dev);
1319         if (ret)
1320                 DRV_LOG(WARNING, "port %u some Verbs Tx queue still remain",
1321                         dev->data->port_id);
1322         ret = mlx5_txq_verify(dev);
1323         if (ret)
1324                 DRV_LOG(WARNING, "port %u some Tx queues still remain",
1325                         dev->data->port_id);
1326         ret = mlx5_flow_verify(dev);
1327         if (ret)
1328                 DRV_LOG(WARNING, "port %u some flows still remain",
1329                         dev->data->port_id);
1330         if (priv->domain_id != RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID) {
1331                 unsigned int c = 0;
1332                 uint16_t port_id;
1333
1334                 MLX5_ETH_FOREACH_DEV(port_id, priv->pci_dev) {
1335                         struct mlx5_priv *opriv =
1336                                 rte_eth_devices[port_id].data->dev_private;
1337
1338                         if (!opriv ||
1339                             opriv->domain_id != priv->domain_id ||
1340                             &rte_eth_devices[port_id] == dev)
1341                                 continue;
1342                         ++c;
1343                         break;
1344                 }
1345                 if (!c)
1346                         claim_zero(rte_eth_switch_domain_free(priv->domain_id));
1347         }
1348         memset(priv, 0, sizeof(*priv));
1349         priv->domain_id = RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID;
1350         /*
1351          * Reset mac_addrs to NULL such that it is not freed as part of
1352          * rte_eth_dev_release_port(). mac_addrs is part of dev_private so
1353          * it is freed when dev_private is freed.
1354          */
1355         dev->data->mac_addrs = NULL;
1356 }
1357
1358 const struct eth_dev_ops mlx5_dev_ops = {
1359         .dev_configure = mlx5_dev_configure,
1360         .dev_start = mlx5_dev_start,
1361         .dev_stop = mlx5_dev_stop,
1362         .dev_set_link_down = mlx5_set_link_down,
1363         .dev_set_link_up = mlx5_set_link_up,
1364         .dev_close = mlx5_dev_close,
1365         .promiscuous_enable = mlx5_promiscuous_enable,
1366         .promiscuous_disable = mlx5_promiscuous_disable,
1367         .allmulticast_enable = mlx5_allmulticast_enable,
1368         .allmulticast_disable = mlx5_allmulticast_disable,
1369         .link_update = mlx5_link_update,
1370         .stats_get = mlx5_stats_get,
1371         .stats_reset = mlx5_stats_reset,
1372         .xstats_get = mlx5_xstats_get,
1373         .xstats_reset = mlx5_xstats_reset,
1374         .xstats_get_names = mlx5_xstats_get_names,
1375         .fw_version_get = mlx5_fw_version_get,
1376         .dev_infos_get = mlx5_dev_infos_get,
1377         .read_clock = mlx5_read_clock,
1378         .dev_supported_ptypes_get = mlx5_dev_supported_ptypes_get,
1379         .vlan_filter_set = mlx5_vlan_filter_set,
1380         .rx_queue_setup = mlx5_rx_queue_setup,
1381         .rx_hairpin_queue_setup = mlx5_rx_hairpin_queue_setup,
1382         .tx_queue_setup = mlx5_tx_queue_setup,
1383         .tx_hairpin_queue_setup = mlx5_tx_hairpin_queue_setup,
1384         .rx_queue_release = mlx5_rx_queue_release,
1385         .tx_queue_release = mlx5_tx_queue_release,
1386         .flow_ctrl_get = mlx5_dev_get_flow_ctrl,
1387         .flow_ctrl_set = mlx5_dev_set_flow_ctrl,
1388         .mac_addr_remove = mlx5_mac_addr_remove,
1389         .mac_addr_add = mlx5_mac_addr_add,
1390         .mac_addr_set = mlx5_mac_addr_set,
1391         .set_mc_addr_list = mlx5_set_mc_addr_list,
1392         .mtu_set = mlx5_dev_set_mtu,
1393         .vlan_strip_queue_set = mlx5_vlan_strip_queue_set,
1394         .vlan_offload_set = mlx5_vlan_offload_set,
1395         .reta_update = mlx5_dev_rss_reta_update,
1396         .reta_query = mlx5_dev_rss_reta_query,
1397         .rss_hash_update = mlx5_rss_hash_update,
1398         .rss_hash_conf_get = mlx5_rss_hash_conf_get,
1399         .filter_ctrl = mlx5_dev_filter_ctrl,
1400         .rx_descriptor_status = mlx5_rx_descriptor_status,
1401         .tx_descriptor_status = mlx5_tx_descriptor_status,
1402         .rxq_info_get = mlx5_rxq_info_get,
1403         .txq_info_get = mlx5_txq_info_get,
1404         .rx_burst_mode_get = mlx5_rx_burst_mode_get,
1405         .tx_burst_mode_get = mlx5_tx_burst_mode_get,
1406         .rx_queue_count = mlx5_rx_queue_count,
1407         .rx_queue_intr_enable = mlx5_rx_intr_enable,
1408         .rx_queue_intr_disable = mlx5_rx_intr_disable,
1409         .is_removed = mlx5_is_removed,
1410         .udp_tunnel_port_add  = mlx5_udp_tunnel_port_add,
1411         .get_module_info = mlx5_get_module_info,
1412         .get_module_eeprom = mlx5_get_module_eeprom,
1413         .hairpin_cap_get = mlx5_hairpin_cap_get,
1414         .mtr_ops_get = mlx5_flow_meter_ops_get,
1415 };
1416
1417 /* Available operations from secondary process. */
1418 static const struct eth_dev_ops mlx5_dev_sec_ops = {
1419         .stats_get = mlx5_stats_get,
1420         .stats_reset = mlx5_stats_reset,
1421         .xstats_get = mlx5_xstats_get,
1422         .xstats_reset = mlx5_xstats_reset,
1423         .xstats_get_names = mlx5_xstats_get_names,
1424         .fw_version_get = mlx5_fw_version_get,
1425         .dev_infos_get = mlx5_dev_infos_get,
1426         .rx_descriptor_status = mlx5_rx_descriptor_status,
1427         .tx_descriptor_status = mlx5_tx_descriptor_status,
1428         .rxq_info_get = mlx5_rxq_info_get,
1429         .txq_info_get = mlx5_txq_info_get,
1430         .rx_burst_mode_get = mlx5_rx_burst_mode_get,
1431         .tx_burst_mode_get = mlx5_tx_burst_mode_get,
1432         .get_module_info = mlx5_get_module_info,
1433         .get_module_eeprom = mlx5_get_module_eeprom,
1434 };
1435
1436 /* Available operations in flow isolated mode. */
1437 const struct eth_dev_ops mlx5_dev_ops_isolate = {
1438         .dev_configure = mlx5_dev_configure,
1439         .dev_start = mlx5_dev_start,
1440         .dev_stop = mlx5_dev_stop,
1441         .dev_set_link_down = mlx5_set_link_down,
1442         .dev_set_link_up = mlx5_set_link_up,
1443         .dev_close = mlx5_dev_close,
1444         .promiscuous_enable = mlx5_promiscuous_enable,
1445         .promiscuous_disable = mlx5_promiscuous_disable,
1446         .allmulticast_enable = mlx5_allmulticast_enable,
1447         .allmulticast_disable = mlx5_allmulticast_disable,
1448         .link_update = mlx5_link_update,
1449         .stats_get = mlx5_stats_get,
1450         .stats_reset = mlx5_stats_reset,
1451         .xstats_get = mlx5_xstats_get,
1452         .xstats_reset = mlx5_xstats_reset,
1453         .xstats_get_names = mlx5_xstats_get_names,
1454         .fw_version_get = mlx5_fw_version_get,
1455         .dev_infos_get = mlx5_dev_infos_get,
1456         .dev_supported_ptypes_get = mlx5_dev_supported_ptypes_get,
1457         .vlan_filter_set = mlx5_vlan_filter_set,
1458         .rx_queue_setup = mlx5_rx_queue_setup,
1459         .rx_hairpin_queue_setup = mlx5_rx_hairpin_queue_setup,
1460         .tx_queue_setup = mlx5_tx_queue_setup,
1461         .tx_hairpin_queue_setup = mlx5_tx_hairpin_queue_setup,
1462         .rx_queue_release = mlx5_rx_queue_release,
1463         .tx_queue_release = mlx5_tx_queue_release,
1464         .flow_ctrl_get = mlx5_dev_get_flow_ctrl,
1465         .flow_ctrl_set = mlx5_dev_set_flow_ctrl,
1466         .mac_addr_remove = mlx5_mac_addr_remove,
1467         .mac_addr_add = mlx5_mac_addr_add,
1468         .mac_addr_set = mlx5_mac_addr_set,
1469         .set_mc_addr_list = mlx5_set_mc_addr_list,
1470         .mtu_set = mlx5_dev_set_mtu,
1471         .vlan_strip_queue_set = mlx5_vlan_strip_queue_set,
1472         .vlan_offload_set = mlx5_vlan_offload_set,
1473         .filter_ctrl = mlx5_dev_filter_ctrl,
1474         .rx_descriptor_status = mlx5_rx_descriptor_status,
1475         .tx_descriptor_status = mlx5_tx_descriptor_status,
1476         .rxq_info_get = mlx5_rxq_info_get,
1477         .txq_info_get = mlx5_txq_info_get,
1478         .rx_burst_mode_get = mlx5_rx_burst_mode_get,
1479         .tx_burst_mode_get = mlx5_tx_burst_mode_get,
1480         .rx_queue_intr_enable = mlx5_rx_intr_enable,
1481         .rx_queue_intr_disable = mlx5_rx_intr_disable,
1482         .is_removed = mlx5_is_removed,
1483         .get_module_info = mlx5_get_module_info,
1484         .get_module_eeprom = mlx5_get_module_eeprom,
1485         .hairpin_cap_get = mlx5_hairpin_cap_get,
1486         .mtr_ops_get = mlx5_flow_meter_ops_get,
1487 };
1488
1489 /**
1490  * Verify and store value for device argument.
1491  *
1492  * @param[in] key
1493  *   Key argument to verify.
1494  * @param[in] val
1495  *   Value associated with key.
1496  * @param opaque
1497  *   User data.
1498  *
1499  * @return
1500  *   0 on success, a negative errno value otherwise and rte_errno is set.
1501  */
1502 static int
1503 mlx5_args_check(const char *key, const char *val, void *opaque)
1504 {
1505         struct mlx5_dev_config *config = opaque;
1506         unsigned long tmp;
1507
1508         /* No-op, port representors are processed in mlx5_dev_spawn(). */
1509         if (!strcmp(MLX5_REPRESENTOR, key))
1510                 return 0;
1511         errno = 0;
1512         tmp = strtoul(val, NULL, 0);
1513         if (errno) {
1514                 rte_errno = errno;
1515                 DRV_LOG(WARNING, "%s: \"%s\" is not a valid integer", key, val);
1516                 return -rte_errno;
1517         }
1518         if (strcmp(MLX5_RXQ_CQE_COMP_EN, key) == 0) {
1519                 config->cqe_comp = !!tmp;
1520         } else if (strcmp(MLX5_RXQ_CQE_PAD_EN, key) == 0) {
1521                 config->cqe_pad = !!tmp;
1522         } else if (strcmp(MLX5_RXQ_PKT_PAD_EN, key) == 0) {
1523                 config->hw_padding = !!tmp;
1524         } else if (strcmp(MLX5_RX_MPRQ_EN, key) == 0) {
1525                 config->mprq.enabled = !!tmp;
1526         } else if (strcmp(MLX5_RX_MPRQ_LOG_STRIDE_NUM, key) == 0) {
1527                 config->mprq.stride_num_n = tmp;
1528         } else if (strcmp(MLX5_RX_MPRQ_MAX_MEMCPY_LEN, key) == 0) {
1529                 config->mprq.max_memcpy_len = tmp;
1530         } else if (strcmp(MLX5_RXQS_MIN_MPRQ, key) == 0) {
1531                 config->mprq.min_rxqs_num = tmp;
1532         } else if (strcmp(MLX5_TXQ_INLINE, key) == 0) {
1533                 DRV_LOG(WARNING, "%s: deprecated parameter,"
1534                                  " converted to txq_inline_max", key);
1535                 config->txq_inline_max = tmp;
1536         } else if (strcmp(MLX5_TXQ_INLINE_MAX, key) == 0) {
1537                 config->txq_inline_max = tmp;
1538         } else if (strcmp(MLX5_TXQ_INLINE_MIN, key) == 0) {
1539                 config->txq_inline_min = tmp;
1540         } else if (strcmp(MLX5_TXQ_INLINE_MPW, key) == 0) {
1541                 config->txq_inline_mpw = tmp;
1542         } else if (strcmp(MLX5_TXQS_MIN_INLINE, key) == 0) {
1543                 config->txqs_inline = tmp;
1544         } else if (strcmp(MLX5_TXQS_MAX_VEC, key) == 0) {
1545                 DRV_LOG(WARNING, "%s: deprecated parameter, ignored", key);
1546         } else if (strcmp(MLX5_TXQ_MPW_EN, key) == 0) {
1547                 config->mps = !!tmp;
1548         } else if (strcmp(MLX5_TX_DB_NC, key) == 0) {
1549                 if (tmp != MLX5_TXDB_CACHED &&
1550                     tmp != MLX5_TXDB_NCACHED &&
1551                     tmp != MLX5_TXDB_HEURISTIC) {
1552                         DRV_LOG(ERR, "invalid Tx doorbell "
1553                                      "mapping parameter");
1554                         rte_errno = EINVAL;
1555                         return -rte_errno;
1556                 }
1557                 config->dbnc = tmp;
1558         } else if (strcmp(MLX5_TXQ_MPW_HDR_DSEG_EN, key) == 0) {
1559                 DRV_LOG(WARNING, "%s: deprecated parameter, ignored", key);
1560         } else if (strcmp(MLX5_TXQ_MAX_INLINE_LEN, key) == 0) {
1561                 DRV_LOG(WARNING, "%s: deprecated parameter,"
1562                                  " converted to txq_inline_mpw", key);
1563                 config->txq_inline_mpw = tmp;
1564         } else if (strcmp(MLX5_TX_VEC_EN, key) == 0) {
1565                 DRV_LOG(WARNING, "%s: deprecated parameter, ignored", key);
1566         } else if (strcmp(MLX5_RX_VEC_EN, key) == 0) {
1567                 config->rx_vec_en = !!tmp;
1568         } else if (strcmp(MLX5_L3_VXLAN_EN, key) == 0) {
1569                 config->l3_vxlan_en = !!tmp;
1570         } else if (strcmp(MLX5_VF_NL_EN, key) == 0) {
1571                 config->vf_nl_en = !!tmp;
1572         } else if (strcmp(MLX5_DV_ESW_EN, key) == 0) {
1573                 config->dv_esw_en = !!tmp;
1574         } else if (strcmp(MLX5_DV_FLOW_EN, key) == 0) {
1575                 config->dv_flow_en = !!tmp;
1576         } else if (strcmp(MLX5_DV_XMETA_EN, key) == 0) {
1577                 if (tmp != MLX5_XMETA_MODE_LEGACY &&
1578                     tmp != MLX5_XMETA_MODE_META16 &&
1579                     tmp != MLX5_XMETA_MODE_META32) {
1580                         DRV_LOG(ERR, "invalid extensive "
1581                                      "metadata parameter");
1582                         rte_errno = EINVAL;
1583                         return -rte_errno;
1584                 }
1585                 config->dv_xmeta_en = tmp;
1586         } else if (strcmp(MLX5_MR_EXT_MEMSEG_EN, key) == 0) {
1587                 config->mr_ext_memseg_en = !!tmp;
1588         } else if (strcmp(MLX5_MAX_DUMP_FILES_NUM, key) == 0) {
1589                 config->max_dump_files_num = tmp;
1590         } else if (strcmp(MLX5_LRO_TIMEOUT_USEC, key) == 0) {
1591                 config->lro.timeout = tmp;
1592         } else if (strcmp(MLX5_CLASS_ARG_NAME, key) == 0) {
1593                 DRV_LOG(DEBUG, "class argument is %s.", val);
1594         } else {
1595                 DRV_LOG(WARNING, "%s: unknown parameter", key);
1596                 rte_errno = EINVAL;
1597                 return -rte_errno;
1598         }
1599         return 0;
1600 }
1601
1602 /**
1603  * Parse device parameters.
1604  *
1605  * @param config
1606  *   Pointer to device configuration structure.
1607  * @param devargs
1608  *   Device arguments structure.
1609  *
1610  * @return
1611  *   0 on success, a negative errno value otherwise and rte_errno is set.
1612  */
1613 static int
1614 mlx5_args(struct mlx5_dev_config *config, struct rte_devargs *devargs)
1615 {
1616         const char **params = (const char *[]){
1617                 MLX5_RXQ_CQE_COMP_EN,
1618                 MLX5_RXQ_CQE_PAD_EN,
1619                 MLX5_RXQ_PKT_PAD_EN,
1620                 MLX5_RX_MPRQ_EN,
1621                 MLX5_RX_MPRQ_LOG_STRIDE_NUM,
1622                 MLX5_RX_MPRQ_MAX_MEMCPY_LEN,
1623                 MLX5_RXQS_MIN_MPRQ,
1624                 MLX5_TXQ_INLINE,
1625                 MLX5_TXQ_INLINE_MIN,
1626                 MLX5_TXQ_INLINE_MAX,
1627                 MLX5_TXQ_INLINE_MPW,
1628                 MLX5_TXQS_MIN_INLINE,
1629                 MLX5_TXQS_MAX_VEC,
1630                 MLX5_TXQ_MPW_EN,
1631                 MLX5_TXQ_MPW_HDR_DSEG_EN,
1632                 MLX5_TXQ_MAX_INLINE_LEN,
1633                 MLX5_TX_DB_NC,
1634                 MLX5_TX_VEC_EN,
1635                 MLX5_RX_VEC_EN,
1636                 MLX5_L3_VXLAN_EN,
1637                 MLX5_VF_NL_EN,
1638                 MLX5_DV_ESW_EN,
1639                 MLX5_DV_FLOW_EN,
1640                 MLX5_DV_XMETA_EN,
1641                 MLX5_MR_EXT_MEMSEG_EN,
1642                 MLX5_REPRESENTOR,
1643                 MLX5_MAX_DUMP_FILES_NUM,
1644                 MLX5_LRO_TIMEOUT_USEC,
1645                 MLX5_CLASS_ARG_NAME,
1646                 NULL,
1647         };
1648         struct rte_kvargs *kvlist;
1649         int ret = 0;
1650         int i;
1651
1652         if (devargs == NULL)
1653                 return 0;
1654         /* Following UGLY cast is done to pass checkpatch. */
1655         kvlist = rte_kvargs_parse(devargs->args, params);
1656         if (kvlist == NULL) {
1657                 rte_errno = EINVAL;
1658                 return -rte_errno;
1659         }
1660         /* Process parameters. */
1661         for (i = 0; (params[i] != NULL); ++i) {
1662                 if (rte_kvargs_count(kvlist, params[i])) {
1663                         ret = rte_kvargs_process(kvlist, params[i],
1664                                                  mlx5_args_check, config);
1665                         if (ret) {
1666                                 rte_errno = EINVAL;
1667                                 rte_kvargs_free(kvlist);
1668                                 return -rte_errno;
1669                         }
1670                 }
1671         }
1672         rte_kvargs_free(kvlist);
1673         return 0;
1674 }
1675
1676 static struct rte_pci_driver mlx5_driver;
1677
1678 /**
1679  * PMD global initialization.
1680  *
1681  * Independent from individual device, this function initializes global
1682  * per-PMD data structures distinguishing primary and secondary processes.
1683  * Hence, each initialization is called once per a process.
1684  *
1685  * @return
1686  *   0 on success, a negative errno value otherwise and rte_errno is set.
1687  */
1688 static int
1689 mlx5_init_once(void)
1690 {
1691         struct mlx5_shared_data *sd;
1692         struct mlx5_local_data *ld = &mlx5_local_data;
1693         int ret = 0;
1694
1695         if (mlx5_init_shared_data())
1696                 return -rte_errno;
1697         sd = mlx5_shared_data;
1698         MLX5_ASSERT(sd);
1699         rte_spinlock_lock(&sd->lock);
1700         switch (rte_eal_process_type()) {
1701         case RTE_PROC_PRIMARY:
1702                 if (sd->init_done)
1703                         break;
1704                 LIST_INIT(&sd->mem_event_cb_list);
1705                 rte_rwlock_init(&sd->mem_event_rwlock);
1706                 rte_mem_event_callback_register("MLX5_MEM_EVENT_CB",
1707                                                 mlx5_mr_mem_event_cb, NULL);
1708                 ret = mlx5_mp_init_primary();
1709                 if (ret)
1710                         goto out;
1711                 sd->init_done = true;
1712                 break;
1713         case RTE_PROC_SECONDARY:
1714                 if (ld->init_done)
1715                         break;
1716                 ret = mlx5_mp_init_secondary();
1717                 if (ret)
1718                         goto out;
1719                 ++sd->secondary_cnt;
1720                 ld->init_done = true;
1721                 break;
1722         default:
1723                 break;
1724         }
1725 out:
1726         rte_spinlock_unlock(&sd->lock);
1727         return ret;
1728 }
1729
1730 /**
1731  * Configures the minimal amount of data to inline into WQE
1732  * while sending packets.
1733  *
1734  * - the txq_inline_min has the maximal priority, if this
1735  *   key is specified in devargs
1736  * - if DevX is enabled the inline mode is queried from the
1737  *   device (HCA attributes and NIC vport context if needed).
1738  * - otherwise L2 mode (18 bytes) is assumed for ConnectX-4/4 Lx
1739  *   and none (0 bytes) for other NICs
1740  *
1741  * @param spawn
1742  *   Verbs device parameters (name, port, switch_info) to spawn.
1743  * @param config
1744  *   Device configuration parameters.
1745  */
1746 static void
1747 mlx5_set_min_inline(struct mlx5_dev_spawn_data *spawn,
1748                     struct mlx5_dev_config *config)
1749 {
1750         if (config->txq_inline_min != MLX5_ARG_UNSET) {
1751                 /* Application defines size of inlined data explicitly. */
1752                 switch (spawn->pci_dev->id.device_id) {
1753                 case PCI_DEVICE_ID_MELLANOX_CONNECTX4:
1754                 case PCI_DEVICE_ID_MELLANOX_CONNECTX4VF:
1755                         if (config->txq_inline_min <
1756                                        (int)MLX5_INLINE_HSIZE_L2) {
1757                                 DRV_LOG(DEBUG,
1758                                         "txq_inline_mix aligned to minimal"
1759                                         " ConnectX-4 required value %d",
1760                                         (int)MLX5_INLINE_HSIZE_L2);
1761                                 config->txq_inline_min = MLX5_INLINE_HSIZE_L2;
1762                         }
1763                         break;
1764                 }
1765                 goto exit;
1766         }
1767         if (config->hca_attr.eth_net_offloads) {
1768                 /* We have DevX enabled, inline mode queried successfully. */
1769                 switch (config->hca_attr.wqe_inline_mode) {
1770                 case MLX5_CAP_INLINE_MODE_L2:
1771                         /* outer L2 header must be inlined. */
1772                         config->txq_inline_min = MLX5_INLINE_HSIZE_L2;
1773                         goto exit;
1774                 case MLX5_CAP_INLINE_MODE_NOT_REQUIRED:
1775                         /* No inline data are required by NIC. */
1776                         config->txq_inline_min = MLX5_INLINE_HSIZE_NONE;
1777                         config->hw_vlan_insert =
1778                                 config->hca_attr.wqe_vlan_insert;
1779                         DRV_LOG(DEBUG, "Tx VLAN insertion is supported");
1780                         goto exit;
1781                 case MLX5_CAP_INLINE_MODE_VPORT_CONTEXT:
1782                         /* inline mode is defined by NIC vport context. */
1783                         if (!config->hca_attr.eth_virt)
1784                                 break;
1785                         switch (config->hca_attr.vport_inline_mode) {
1786                         case MLX5_INLINE_MODE_NONE:
1787                                 config->txq_inline_min =
1788                                         MLX5_INLINE_HSIZE_NONE;
1789                                 goto exit;
1790                         case MLX5_INLINE_MODE_L2:
1791                                 config->txq_inline_min =
1792                                         MLX5_INLINE_HSIZE_L2;
1793                                 goto exit;
1794                         case MLX5_INLINE_MODE_IP:
1795                                 config->txq_inline_min =
1796                                         MLX5_INLINE_HSIZE_L3;
1797                                 goto exit;
1798                         case MLX5_INLINE_MODE_TCP_UDP:
1799                                 config->txq_inline_min =
1800                                         MLX5_INLINE_HSIZE_L4;
1801                                 goto exit;
1802                         case MLX5_INLINE_MODE_INNER_L2:
1803                                 config->txq_inline_min =
1804                                         MLX5_INLINE_HSIZE_INNER_L2;
1805                                 goto exit;
1806                         case MLX5_INLINE_MODE_INNER_IP:
1807                                 config->txq_inline_min =
1808                                         MLX5_INLINE_HSIZE_INNER_L3;
1809                                 goto exit;
1810                         case MLX5_INLINE_MODE_INNER_TCP_UDP:
1811                                 config->txq_inline_min =
1812                                         MLX5_INLINE_HSIZE_INNER_L4;
1813                                 goto exit;
1814                         }
1815                 }
1816         }
1817         /*
1818          * We get here if we are unable to deduce
1819          * inline data size with DevX. Try PCI ID
1820          * to determine old NICs.
1821          */
1822         switch (spawn->pci_dev->id.device_id) {
1823         case PCI_DEVICE_ID_MELLANOX_CONNECTX4:
1824         case PCI_DEVICE_ID_MELLANOX_CONNECTX4VF:
1825         case PCI_DEVICE_ID_MELLANOX_CONNECTX4LX:
1826         case PCI_DEVICE_ID_MELLANOX_CONNECTX4LXVF:
1827                 config->txq_inline_min = MLX5_INLINE_HSIZE_L2;
1828                 config->hw_vlan_insert = 0;
1829                 break;
1830         case PCI_DEVICE_ID_MELLANOX_CONNECTX5:
1831         case PCI_DEVICE_ID_MELLANOX_CONNECTX5VF:
1832         case PCI_DEVICE_ID_MELLANOX_CONNECTX5EX:
1833         case PCI_DEVICE_ID_MELLANOX_CONNECTX5EXVF:
1834                 /*
1835                  * These NICs support VLAN insertion from WQE and
1836                  * report the wqe_vlan_insert flag. But there is the bug
1837                  * and PFC control may be broken, so disable feature.
1838                  */
1839                 config->hw_vlan_insert = 0;
1840                 config->txq_inline_min = MLX5_INLINE_HSIZE_NONE;
1841                 break;
1842         default:
1843                 config->txq_inline_min = MLX5_INLINE_HSIZE_NONE;
1844                 break;
1845         }
1846 exit:
1847         DRV_LOG(DEBUG, "min tx inline configured: %d", config->txq_inline_min);
1848 }
1849
1850 /**
1851  * Configures the metadata mask fields in the shared context.
1852  *
1853  * @param [in] dev
1854  *   Pointer to Ethernet device.
1855  */
1856 static void
1857 mlx5_set_metadata_mask(struct rte_eth_dev *dev)
1858 {
1859         struct mlx5_priv *priv = dev->data->dev_private;
1860         struct mlx5_ibv_shared *sh = priv->sh;
1861         uint32_t meta, mark, reg_c0;
1862
1863         reg_c0 = ~priv->vport_meta_mask;
1864         switch (priv->config.dv_xmeta_en) {
1865         case MLX5_XMETA_MODE_LEGACY:
1866                 meta = UINT32_MAX;
1867                 mark = MLX5_FLOW_MARK_MASK;
1868                 break;
1869         case MLX5_XMETA_MODE_META16:
1870                 meta = reg_c0 >> rte_bsf32(reg_c0);
1871                 mark = MLX5_FLOW_MARK_MASK;
1872                 break;
1873         case MLX5_XMETA_MODE_META32:
1874                 meta = UINT32_MAX;
1875                 mark = (reg_c0 >> rte_bsf32(reg_c0)) & MLX5_FLOW_MARK_MASK;
1876                 break;
1877         default:
1878                 meta = 0;
1879                 mark = 0;
1880                 MLX5_ASSERT(false);
1881                 break;
1882         }
1883         if (sh->dv_mark_mask && sh->dv_mark_mask != mark)
1884                 DRV_LOG(WARNING, "metadata MARK mask mismatche %08X:%08X",
1885                                  sh->dv_mark_mask, mark);
1886         else
1887                 sh->dv_mark_mask = mark;
1888         if (sh->dv_meta_mask && sh->dv_meta_mask != meta)
1889                 DRV_LOG(WARNING, "metadata META mask mismatche %08X:%08X",
1890                                  sh->dv_meta_mask, meta);
1891         else
1892                 sh->dv_meta_mask = meta;
1893         if (sh->dv_regc0_mask && sh->dv_regc0_mask != reg_c0)
1894                 DRV_LOG(WARNING, "metadata reg_c0 mask mismatche %08X:%08X",
1895                                  sh->dv_meta_mask, reg_c0);
1896         else
1897                 sh->dv_regc0_mask = reg_c0;
1898         DRV_LOG(DEBUG, "metadata mode %u", priv->config.dv_xmeta_en);
1899         DRV_LOG(DEBUG, "metadata MARK mask %08X", sh->dv_mark_mask);
1900         DRV_LOG(DEBUG, "metadata META mask %08X", sh->dv_meta_mask);
1901         DRV_LOG(DEBUG, "metadata reg_c0 mask %08X", sh->dv_regc0_mask);
1902 }
1903
1904 /**
1905  * Allocate page of door-bells and register it using DevX API.
1906  *
1907  * @param [in] dev
1908  *   Pointer to Ethernet device.
1909  *
1910  * @return
1911  *   Pointer to new page on success, NULL otherwise.
1912  */
1913 static struct mlx5_devx_dbr_page *
1914 mlx5_alloc_dbr_page(struct rte_eth_dev *dev)
1915 {
1916         struct mlx5_priv *priv = dev->data->dev_private;
1917         struct mlx5_devx_dbr_page *page;
1918
1919         /* Allocate space for door-bell page and management data. */
1920         page = rte_calloc_socket(__func__, 1, sizeof(struct mlx5_devx_dbr_page),
1921                                  RTE_CACHE_LINE_SIZE, dev->device->numa_node);
1922         if (!page) {
1923                 DRV_LOG(ERR, "port %u cannot allocate dbr page",
1924                         dev->data->port_id);
1925                 return NULL;
1926         }
1927         /* Register allocated memory. */
1928         page->umem = mlx5_glue->devx_umem_reg(priv->sh->ctx, page->dbrs,
1929                                               MLX5_DBR_PAGE_SIZE, 0);
1930         if (!page->umem) {
1931                 DRV_LOG(ERR, "port %u cannot umem reg dbr page",
1932                         dev->data->port_id);
1933                 rte_free(page);
1934                 return NULL;
1935         }
1936         return page;
1937 }
1938
1939 /**
1940  * Find the next available door-bell, allocate new page if needed.
1941  *
1942  * @param [in] dev
1943  *   Pointer to Ethernet device.
1944  * @param [out] dbr_page
1945  *   Door-bell page containing the page data.
1946  *
1947  * @return
1948  *   Door-bell address offset on success, a negative error value otherwise.
1949  */
1950 int64_t
1951 mlx5_get_dbr(struct rte_eth_dev *dev, struct mlx5_devx_dbr_page **dbr_page)
1952 {
1953         struct mlx5_priv *priv = dev->data->dev_private;
1954         struct mlx5_devx_dbr_page *page = NULL;
1955         uint32_t i, j;
1956
1957         LIST_FOREACH(page, &priv->dbrpgs, next)
1958                 if (page->dbr_count < MLX5_DBR_PER_PAGE)
1959                         break;
1960         if (!page) { /* No page with free door-bell exists. */
1961                 page = mlx5_alloc_dbr_page(dev);
1962                 if (!page) /* Failed to allocate new page. */
1963                         return (-1);
1964                 LIST_INSERT_HEAD(&priv->dbrpgs, page, next);
1965         }
1966         /* Loop to find bitmap part with clear bit. */
1967         for (i = 0;
1968              i < MLX5_DBR_BITMAP_SIZE && page->dbr_bitmap[i] == UINT64_MAX;
1969              i++)
1970                 ; /* Empty. */
1971         /* Find the first clear bit. */
1972         j = rte_bsf64(~page->dbr_bitmap[i]);
1973         MLX5_ASSERT(i < (MLX5_DBR_PER_PAGE / 64));
1974         page->dbr_bitmap[i] |= (1 << j);
1975         page->dbr_count++;
1976         *dbr_page = page;
1977         return (((i * 64) + j) * sizeof(uint64_t));
1978 }
1979
1980 /**
1981  * Release a door-bell record.
1982  *
1983  * @param [in] dev
1984  *   Pointer to Ethernet device.
1985  * @param [in] umem_id
1986  *   UMEM ID of page containing the door-bell record to release.
1987  * @param [in] offset
1988  *   Offset of door-bell record in page.
1989  *
1990  * @return
1991  *   0 on success, a negative error value otherwise.
1992  */
1993 int32_t
1994 mlx5_release_dbr(struct rte_eth_dev *dev, uint32_t umem_id, uint64_t offset)
1995 {
1996         struct mlx5_priv *priv = dev->data->dev_private;
1997         struct mlx5_devx_dbr_page *page = NULL;
1998         int ret = 0;
1999
2000         LIST_FOREACH(page, &priv->dbrpgs, next)
2001                 /* Find the page this address belongs to. */
2002                 if (page->umem->umem_id == umem_id)
2003                         break;
2004         if (!page)
2005                 return -EINVAL;
2006         page->dbr_count--;
2007         if (!page->dbr_count) {
2008                 /* Page not used, free it and remove from list. */
2009                 LIST_REMOVE(page, next);
2010                 if (page->umem)
2011                         ret = -mlx5_glue->devx_umem_dereg(page->umem);
2012                 rte_free(page);
2013         } else {
2014                 /* Mark in bitmap that this door-bell is not in use. */
2015                 offset /= MLX5_DBR_SIZE;
2016                 int i = offset / 64;
2017                 int j = offset % 64;
2018
2019                 page->dbr_bitmap[i] &= ~(1 << j);
2020         }
2021         return ret;
2022 }
2023
2024 int
2025 rte_pmd_mlx5_get_dyn_flag_names(char *names[], unsigned int n)
2026 {
2027         static const char *const dynf_names[] = {
2028                 RTE_PMD_MLX5_FINE_GRANULARITY_INLINE,
2029                 RTE_MBUF_DYNFLAG_METADATA_NAME
2030         };
2031         unsigned int i;
2032
2033         if (n < RTE_DIM(dynf_names))
2034                 return -ENOMEM;
2035         for (i = 0; i < RTE_DIM(dynf_names); i++) {
2036                 if (names[i] == NULL)
2037                         return -EINVAL;
2038                 strcpy(names[i], dynf_names[i]);
2039         }
2040         return RTE_DIM(dynf_names);
2041 }
2042
2043 /**
2044  * Check sibling device configurations.
2045  *
2046  * Sibling devices sharing the Infiniband device context
2047  * should have compatible configurations. This regards
2048  * representors and bonding slaves.
2049  *
2050  * @param priv
2051  *   Private device descriptor.
2052  * @param config
2053  *   Configuration of the device is going to be created.
2054  *
2055  * @return
2056  *   0 on success, EINVAL otherwise
2057  */
2058 static int
2059 mlx5_dev_check_sibling_config(struct mlx5_priv *priv,
2060                               struct mlx5_dev_config *config)
2061 {
2062         struct mlx5_ibv_shared *sh = priv->sh;
2063         struct mlx5_dev_config *sh_conf = NULL;
2064         uint16_t port_id;
2065
2066         MLX5_ASSERT(sh);
2067         /* Nothing to compare for the single/first device. */
2068         if (sh->refcnt == 1)
2069                 return 0;
2070         /* Find the device with shared context. */
2071         MLX5_ETH_FOREACH_DEV(port_id, priv->pci_dev) {
2072                 struct mlx5_priv *opriv =
2073                         rte_eth_devices[port_id].data->dev_private;
2074
2075                 if (opriv && opriv != priv && opriv->sh == sh) {
2076                         sh_conf = &opriv->config;
2077                         break;
2078                 }
2079         }
2080         if (!sh_conf)
2081                 return 0;
2082         if (sh_conf->dv_flow_en ^ config->dv_flow_en) {
2083                 DRV_LOG(ERR, "\"dv_flow_en\" configuration mismatch"
2084                              " for shared %s context", sh->ibdev_name);
2085                 rte_errno = EINVAL;
2086                 return rte_errno;
2087         }
2088         if (sh_conf->dv_xmeta_en ^ config->dv_xmeta_en) {
2089                 DRV_LOG(ERR, "\"dv_xmeta_en\" configuration mismatch"
2090                              " for shared %s context", sh->ibdev_name);
2091                 rte_errno = EINVAL;
2092                 return rte_errno;
2093         }
2094         return 0;
2095 }
2096 /**
2097  * Spawn an Ethernet device from Verbs information.
2098  *
2099  * @param dpdk_dev
2100  *   Backing DPDK device.
2101  * @param spawn
2102  *   Verbs device parameters (name, port, switch_info) to spawn.
2103  * @param config
2104  *   Device configuration parameters.
2105  *
2106  * @return
2107  *   A valid Ethernet device object on success, NULL otherwise and rte_errno
2108  *   is set. The following errors are defined:
2109  *
2110  *   EBUSY: device is not supposed to be spawned.
2111  *   EEXIST: device is already spawned
2112  */
2113 static struct rte_eth_dev *
2114 mlx5_dev_spawn(struct rte_device *dpdk_dev,
2115                struct mlx5_dev_spawn_data *spawn,
2116                struct mlx5_dev_config config)
2117 {
2118         const struct mlx5_switch_info *switch_info = &spawn->info;
2119         struct mlx5_ibv_shared *sh = NULL;
2120         struct ibv_port_attr port_attr;
2121         struct mlx5dv_context dv_attr = { .comp_mask = 0 };
2122         struct rte_eth_dev *eth_dev = NULL;
2123         struct mlx5_priv *priv = NULL;
2124         int err = 0;
2125         unsigned int hw_padding = 0;
2126         unsigned int mps;
2127         unsigned int cqe_comp;
2128         unsigned int cqe_pad = 0;
2129         unsigned int tunnel_en = 0;
2130         unsigned int mpls_en = 0;
2131         unsigned int swp = 0;
2132         unsigned int mprq = 0;
2133         unsigned int mprq_min_stride_size_n = 0;
2134         unsigned int mprq_max_stride_size_n = 0;
2135         unsigned int mprq_min_stride_num_n = 0;
2136         unsigned int mprq_max_stride_num_n = 0;
2137         struct rte_ether_addr mac;
2138         char name[RTE_ETH_NAME_MAX_LEN];
2139         int own_domain_id = 0;
2140         uint16_t port_id;
2141         unsigned int i;
2142 #ifdef HAVE_MLX5DV_DR_DEVX_PORT
2143         struct mlx5dv_devx_port devx_port = { .comp_mask = 0 };
2144 #endif
2145
2146         /* Determine if this port representor is supposed to be spawned. */
2147         if (switch_info->representor && dpdk_dev->devargs) {
2148                 struct rte_eth_devargs eth_da;
2149
2150                 err = rte_eth_devargs_parse(dpdk_dev->devargs->args, &eth_da);
2151                 if (err) {
2152                         rte_errno = -err;
2153                         DRV_LOG(ERR, "failed to process device arguments: %s",
2154                                 strerror(rte_errno));
2155                         return NULL;
2156                 }
2157                 for (i = 0; i < eth_da.nb_representor_ports; ++i)
2158                         if (eth_da.representor_ports[i] ==
2159                             (uint16_t)switch_info->port_name)
2160                                 break;
2161                 if (i == eth_da.nb_representor_ports) {
2162                         rte_errno = EBUSY;
2163                         return NULL;
2164                 }
2165         }
2166         /* Build device name. */
2167         if (spawn->pf_bond <  0) {
2168                 /* Single device. */
2169                 if (!switch_info->representor)
2170                         strlcpy(name, dpdk_dev->name, sizeof(name));
2171                 else
2172                         snprintf(name, sizeof(name), "%s_representor_%u",
2173                                  dpdk_dev->name, switch_info->port_name);
2174         } else {
2175                 /* Bonding device. */
2176                 if (!switch_info->representor)
2177                         snprintf(name, sizeof(name), "%s_%s",
2178                                  dpdk_dev->name, spawn->ibv_dev->name);
2179                 else
2180                         snprintf(name, sizeof(name), "%s_%s_representor_%u",
2181                                  dpdk_dev->name, spawn->ibv_dev->name,
2182                                  switch_info->port_name);
2183         }
2184         /* check if the device is already spawned */
2185         if (rte_eth_dev_get_port_by_name(name, &port_id) == 0) {
2186                 rte_errno = EEXIST;
2187                 return NULL;
2188         }
2189         DRV_LOG(DEBUG, "naming Ethernet device \"%s\"", name);
2190         if (rte_eal_process_type() == RTE_PROC_SECONDARY) {
2191                 eth_dev = rte_eth_dev_attach_secondary(name);
2192                 if (eth_dev == NULL) {
2193                         DRV_LOG(ERR, "can not attach rte ethdev");
2194                         rte_errno = ENOMEM;
2195                         return NULL;
2196                 }
2197                 eth_dev->device = dpdk_dev;
2198                 eth_dev->dev_ops = &mlx5_dev_sec_ops;
2199                 err = mlx5_proc_priv_init(eth_dev);
2200                 if (err)
2201                         return NULL;
2202                 /* Receive command fd from primary process */
2203                 err = mlx5_mp_req_verbs_cmd_fd(eth_dev);
2204                 if (err < 0)
2205                         return NULL;
2206                 /* Remap UAR for Tx queues. */
2207                 err = mlx5_tx_uar_init_secondary(eth_dev, err);
2208                 if (err)
2209                         return NULL;
2210                 /*
2211                  * Ethdev pointer is still required as input since
2212                  * the primary device is not accessible from the
2213                  * secondary process.
2214                  */
2215                 eth_dev->rx_pkt_burst = mlx5_select_rx_function(eth_dev);
2216                 eth_dev->tx_pkt_burst = mlx5_select_tx_function(eth_dev);
2217                 return eth_dev;
2218         }
2219         /*
2220          * Some parameters ("tx_db_nc" in particularly) are needed in
2221          * advance to create dv/verbs device context. We proceed the
2222          * devargs here to get ones, and later proceed devargs again
2223          * to override some hardware settings.
2224          */
2225         err = mlx5_args(&config, dpdk_dev->devargs);
2226         if (err) {
2227                 err = rte_errno;
2228                 DRV_LOG(ERR, "failed to process device arguments: %s",
2229                         strerror(rte_errno));
2230                 goto error;
2231         }
2232         sh = mlx5_alloc_shared_ibctx(spawn, &config);
2233         if (!sh)
2234                 return NULL;
2235         config.devx = sh->devx;
2236 #ifdef HAVE_MLX5DV_DR_ACTION_DEST_DEVX_TIR
2237         config.dest_tir = 1;
2238 #endif
2239 #ifdef HAVE_IBV_MLX5_MOD_SWP
2240         dv_attr.comp_mask |= MLX5DV_CONTEXT_MASK_SWP;
2241 #endif
2242         /*
2243          * Multi-packet send is supported by ConnectX-4 Lx PF as well
2244          * as all ConnectX-5 devices.
2245          */
2246 #ifdef HAVE_IBV_DEVICE_TUNNEL_SUPPORT
2247         dv_attr.comp_mask |= MLX5DV_CONTEXT_MASK_TUNNEL_OFFLOADS;
2248 #endif
2249 #ifdef HAVE_IBV_DEVICE_STRIDING_RQ_SUPPORT
2250         dv_attr.comp_mask |= MLX5DV_CONTEXT_MASK_STRIDING_RQ;
2251 #endif
2252         mlx5_glue->dv_query_device(sh->ctx, &dv_attr);
2253         if (dv_attr.flags & MLX5DV_CONTEXT_FLAGS_MPW_ALLOWED) {
2254                 if (dv_attr.flags & MLX5DV_CONTEXT_FLAGS_ENHANCED_MPW) {
2255                         DRV_LOG(DEBUG, "enhanced MPW is supported");
2256                         mps = MLX5_MPW_ENHANCED;
2257                 } else {
2258                         DRV_LOG(DEBUG, "MPW is supported");
2259                         mps = MLX5_MPW;
2260                 }
2261         } else {
2262                 DRV_LOG(DEBUG, "MPW isn't supported");
2263                 mps = MLX5_MPW_DISABLED;
2264         }
2265 #ifdef HAVE_IBV_MLX5_MOD_SWP
2266         if (dv_attr.comp_mask & MLX5DV_CONTEXT_MASK_SWP)
2267                 swp = dv_attr.sw_parsing_caps.sw_parsing_offloads;
2268         DRV_LOG(DEBUG, "SWP support: %u", swp);
2269 #endif
2270         config.swp = !!swp;
2271 #ifdef HAVE_IBV_DEVICE_STRIDING_RQ_SUPPORT
2272         if (dv_attr.comp_mask & MLX5DV_CONTEXT_MASK_STRIDING_RQ) {
2273                 struct mlx5dv_striding_rq_caps mprq_caps =
2274                         dv_attr.striding_rq_caps;
2275
2276                 DRV_LOG(DEBUG, "\tmin_single_stride_log_num_of_bytes: %d",
2277                         mprq_caps.min_single_stride_log_num_of_bytes);
2278                 DRV_LOG(DEBUG, "\tmax_single_stride_log_num_of_bytes: %d",
2279                         mprq_caps.max_single_stride_log_num_of_bytes);
2280                 DRV_LOG(DEBUG, "\tmin_single_wqe_log_num_of_strides: %d",
2281                         mprq_caps.min_single_wqe_log_num_of_strides);
2282                 DRV_LOG(DEBUG, "\tmax_single_wqe_log_num_of_strides: %d",
2283                         mprq_caps.max_single_wqe_log_num_of_strides);
2284                 DRV_LOG(DEBUG, "\tsupported_qpts: %d",
2285                         mprq_caps.supported_qpts);
2286                 DRV_LOG(DEBUG, "device supports Multi-Packet RQ");
2287                 mprq = 1;
2288                 mprq_min_stride_size_n =
2289                         mprq_caps.min_single_stride_log_num_of_bytes;
2290                 mprq_max_stride_size_n =
2291                         mprq_caps.max_single_stride_log_num_of_bytes;
2292                 mprq_min_stride_num_n =
2293                         mprq_caps.min_single_wqe_log_num_of_strides;
2294                 mprq_max_stride_num_n =
2295                         mprq_caps.max_single_wqe_log_num_of_strides;
2296                 config.mprq.stride_num_n = RTE_MAX(MLX5_MPRQ_STRIDE_NUM_N,
2297                                                    mprq_min_stride_num_n);
2298         }
2299 #endif
2300         if (RTE_CACHE_LINE_SIZE == 128 &&
2301             !(dv_attr.flags & MLX5DV_CONTEXT_FLAGS_CQE_128B_COMP))
2302                 cqe_comp = 0;
2303         else
2304                 cqe_comp = 1;
2305         config.cqe_comp = cqe_comp;
2306 #ifdef HAVE_IBV_MLX5_MOD_CQE_128B_PAD
2307         /* Whether device supports 128B Rx CQE padding. */
2308         cqe_pad = RTE_CACHE_LINE_SIZE == 128 &&
2309                   (dv_attr.flags & MLX5DV_CONTEXT_FLAGS_CQE_128B_PAD);
2310 #endif
2311 #ifdef HAVE_IBV_DEVICE_TUNNEL_SUPPORT
2312         if (dv_attr.comp_mask & MLX5DV_CONTEXT_MASK_TUNNEL_OFFLOADS) {
2313                 tunnel_en = ((dv_attr.tunnel_offloads_caps &
2314                               MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_VXLAN) &&
2315                              (dv_attr.tunnel_offloads_caps &
2316                               MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_GRE) &&
2317                              (dv_attr.tunnel_offloads_caps &
2318                               MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_GENEVE));
2319         }
2320         DRV_LOG(DEBUG, "tunnel offloading is %ssupported",
2321                 tunnel_en ? "" : "not ");
2322 #else
2323         DRV_LOG(WARNING,
2324                 "tunnel offloading disabled due to old OFED/rdma-core version");
2325 #endif
2326         config.tunnel_en = tunnel_en;
2327 #ifdef HAVE_IBV_DEVICE_MPLS_SUPPORT
2328         mpls_en = ((dv_attr.tunnel_offloads_caps &
2329                     MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_CW_MPLS_OVER_GRE) &&
2330                    (dv_attr.tunnel_offloads_caps &
2331                     MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_CW_MPLS_OVER_UDP));
2332         DRV_LOG(DEBUG, "MPLS over GRE/UDP tunnel offloading is %ssupported",
2333                 mpls_en ? "" : "not ");
2334 #else
2335         DRV_LOG(WARNING, "MPLS over GRE/UDP tunnel offloading disabled due to"
2336                 " old OFED/rdma-core version or firmware configuration");
2337 #endif
2338         config.mpls_en = mpls_en;
2339         /* Check port status. */
2340         err = mlx5_glue->query_port(sh->ctx, spawn->ibv_port, &port_attr);
2341         if (err) {
2342                 DRV_LOG(ERR, "port query failed: %s", strerror(err));
2343                 goto error;
2344         }
2345         if (port_attr.link_layer != IBV_LINK_LAYER_ETHERNET) {
2346                 DRV_LOG(ERR, "port is not configured in Ethernet mode");
2347                 err = EINVAL;
2348                 goto error;
2349         }
2350         if (port_attr.state != IBV_PORT_ACTIVE)
2351                 DRV_LOG(DEBUG, "port is not active: \"%s\" (%d)",
2352                         mlx5_glue->port_state_str(port_attr.state),
2353                         port_attr.state);
2354         /* Allocate private eth device data. */
2355         priv = rte_zmalloc("ethdev private structure",
2356                            sizeof(*priv),
2357                            RTE_CACHE_LINE_SIZE);
2358         if (priv == NULL) {
2359                 DRV_LOG(ERR, "priv allocation failure");
2360                 err = ENOMEM;
2361                 goto error;
2362         }
2363         priv->sh = sh;
2364         priv->ibv_port = spawn->ibv_port;
2365         priv->pci_dev = spawn->pci_dev;
2366         priv->mtu = RTE_ETHER_MTU;
2367 #ifndef RTE_ARCH_64
2368         /* Initialize UAR access locks for 32bit implementations. */
2369         rte_spinlock_init(&priv->uar_lock_cq);
2370         for (i = 0; i < MLX5_UAR_PAGE_NUM_MAX; i++)
2371                 rte_spinlock_init(&priv->uar_lock[i]);
2372 #endif
2373         /* Some internal functions rely on Netlink sockets, open them now. */
2374         priv->nl_socket_rdma = mlx5_nl_init(NETLINK_RDMA);
2375         priv->nl_socket_route = mlx5_nl_init(NETLINK_ROUTE);
2376         priv->representor = !!switch_info->representor;
2377         priv->master = !!switch_info->master;
2378         priv->domain_id = RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID;
2379         priv->vport_meta_tag = 0;
2380         priv->vport_meta_mask = 0;
2381         priv->pf_bond = spawn->pf_bond;
2382 #ifdef HAVE_MLX5DV_DR_DEVX_PORT
2383         /*
2384          * The DevX port query API is implemented. E-Switch may use
2385          * either vport or reg_c[0] metadata register to match on
2386          * vport index. The engaged part of metadata register is
2387          * defined by mask.
2388          */
2389         if (switch_info->representor || switch_info->master) {
2390                 devx_port.comp_mask = MLX5DV_DEVX_PORT_VPORT |
2391                                       MLX5DV_DEVX_PORT_MATCH_REG_C_0;
2392                 err = mlx5_glue->devx_port_query(sh->ctx, spawn->ibv_port,
2393                                                  &devx_port);
2394                 if (err) {
2395                         DRV_LOG(WARNING,
2396                                 "can't query devx port %d on device %s",
2397                                 spawn->ibv_port, spawn->ibv_dev->name);
2398                         devx_port.comp_mask = 0;
2399                 }
2400         }
2401         if (devx_port.comp_mask & MLX5DV_DEVX_PORT_MATCH_REG_C_0) {
2402                 priv->vport_meta_tag = devx_port.reg_c_0.value;
2403                 priv->vport_meta_mask = devx_port.reg_c_0.mask;
2404                 if (!priv->vport_meta_mask) {
2405                         DRV_LOG(ERR, "vport zero mask for port %d"
2406                                      " on bonding device %s",
2407                                      spawn->ibv_port, spawn->ibv_dev->name);
2408                         err = ENOTSUP;
2409                         goto error;
2410                 }
2411                 if (priv->vport_meta_tag & ~priv->vport_meta_mask) {
2412                         DRV_LOG(ERR, "invalid vport tag for port %d"
2413                                      " on bonding device %s",
2414                                      spawn->ibv_port, spawn->ibv_dev->name);
2415                         err = ENOTSUP;
2416                         goto error;
2417                 }
2418         }
2419         if (devx_port.comp_mask & MLX5DV_DEVX_PORT_VPORT) {
2420                 priv->vport_id = devx_port.vport_num;
2421         } else if (spawn->pf_bond >= 0) {
2422                 DRV_LOG(ERR, "can't deduce vport index for port %d"
2423                              " on bonding device %s",
2424                              spawn->ibv_port, spawn->ibv_dev->name);
2425                 err = ENOTSUP;
2426                 goto error;
2427         } else {
2428                 /* Suppose vport index in compatible way. */
2429                 priv->vport_id = switch_info->representor ?
2430                                  switch_info->port_name + 1 : -1;
2431         }
2432 #else
2433         /*
2434          * Kernel/rdma_core support single E-Switch per PF configurations
2435          * only and vport_id field contains the vport index for
2436          * associated VF, which is deduced from representor port name.
2437          * For example, let's have the IB device port 10, it has
2438          * attached network device eth0, which has port name attribute
2439          * pf0vf2, we can deduce the VF number as 2, and set vport index
2440          * as 3 (2+1). This assigning schema should be changed if the
2441          * multiple E-Switch instances per PF configurations or/and PCI
2442          * subfunctions are added.
2443          */
2444         priv->vport_id = switch_info->representor ?
2445                          switch_info->port_name + 1 : -1;
2446 #endif
2447         /* representor_id field keeps the unmodified VF index. */
2448         priv->representor_id = switch_info->representor ?
2449                                switch_info->port_name : -1;
2450         /*
2451          * Look for sibling devices in order to reuse their switch domain
2452          * if any, otherwise allocate one.
2453          */
2454         MLX5_ETH_FOREACH_DEV(port_id, priv->pci_dev) {
2455                 const struct mlx5_priv *opriv =
2456                         rte_eth_devices[port_id].data->dev_private;
2457
2458                 if (!opriv ||
2459                     opriv->sh != priv->sh ||
2460                         opriv->domain_id ==
2461                         RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID)
2462                         continue;
2463                 priv->domain_id = opriv->domain_id;
2464                 break;
2465         }
2466         if (priv->domain_id == RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID) {
2467                 err = rte_eth_switch_domain_alloc(&priv->domain_id);
2468                 if (err) {
2469                         err = rte_errno;
2470                         DRV_LOG(ERR, "unable to allocate switch domain: %s",
2471                                 strerror(rte_errno));
2472                         goto error;
2473                 }
2474                 own_domain_id = 1;
2475         }
2476         /* Override some values set by hardware configuration. */
2477         mlx5_args(&config, dpdk_dev->devargs);
2478         err = mlx5_dev_check_sibling_config(priv, &config);
2479         if (err)
2480                 goto error;
2481         config.hw_csum = !!(sh->device_attr.device_cap_flags_ex &
2482                             IBV_DEVICE_RAW_IP_CSUM);
2483         DRV_LOG(DEBUG, "checksum offloading is %ssupported",
2484                 (config.hw_csum ? "" : "not "));
2485 #if !defined(HAVE_IBV_DEVICE_COUNTERS_SET_V42) && \
2486         !defined(HAVE_IBV_DEVICE_COUNTERS_SET_V45)
2487         DRV_LOG(DEBUG, "counters are not supported");
2488 #endif
2489 #if !defined(HAVE_IBV_FLOW_DV_SUPPORT) || !defined(HAVE_MLX5DV_DR)
2490         if (config.dv_flow_en) {
2491                 DRV_LOG(WARNING, "DV flow is not supported");
2492                 config.dv_flow_en = 0;
2493         }
2494 #endif
2495         config.ind_table_max_size =
2496                 sh->device_attr.rss_caps.max_rwq_indirection_table_size;
2497         /*
2498          * Remove this check once DPDK supports larger/variable
2499          * indirection tables.
2500          */
2501         if (config.ind_table_max_size > (unsigned int)ETH_RSS_RETA_SIZE_512)
2502                 config.ind_table_max_size = ETH_RSS_RETA_SIZE_512;
2503         DRV_LOG(DEBUG, "maximum Rx indirection table size is %u",
2504                 config.ind_table_max_size);
2505         config.hw_vlan_strip = !!(sh->device_attr.raw_packet_caps &
2506                                   IBV_RAW_PACKET_CAP_CVLAN_STRIPPING);
2507         DRV_LOG(DEBUG, "VLAN stripping is %ssupported",
2508                 (config.hw_vlan_strip ? "" : "not "));
2509         config.hw_fcs_strip = !!(sh->device_attr.raw_packet_caps &
2510                                  IBV_RAW_PACKET_CAP_SCATTER_FCS);
2511         DRV_LOG(DEBUG, "FCS stripping configuration is %ssupported",
2512                 (config.hw_fcs_strip ? "" : "not "));
2513 #if defined(HAVE_IBV_WQ_FLAG_RX_END_PADDING)
2514         hw_padding = !!sh->device_attr.rx_pad_end_addr_align;
2515 #elif defined(HAVE_IBV_WQ_FLAGS_PCI_WRITE_END_PADDING)
2516         hw_padding = !!(sh->device_attr.device_cap_flags_ex &
2517                         IBV_DEVICE_PCI_WRITE_END_PADDING);
2518 #endif
2519         if (config.hw_padding && !hw_padding) {
2520                 DRV_LOG(DEBUG, "Rx end alignment padding isn't supported");
2521                 config.hw_padding = 0;
2522         } else if (config.hw_padding) {
2523                 DRV_LOG(DEBUG, "Rx end alignment padding is enabled");
2524         }
2525         config.tso = (sh->device_attr.tso_caps.max_tso > 0 &&
2526                       (sh->device_attr.tso_caps.supported_qpts &
2527                        (1 << IBV_QPT_RAW_PACKET)));
2528         if (config.tso)
2529                 config.tso_max_payload_sz = sh->device_attr.tso_caps.max_tso;
2530         /*
2531          * MPW is disabled by default, while the Enhanced MPW is enabled
2532          * by default.
2533          */
2534         if (config.mps == MLX5_ARG_UNSET)
2535                 config.mps = (mps == MLX5_MPW_ENHANCED) ? MLX5_MPW_ENHANCED :
2536                                                           MLX5_MPW_DISABLED;
2537         else
2538                 config.mps = config.mps ? mps : MLX5_MPW_DISABLED;
2539         DRV_LOG(INFO, "%sMPS is %s",
2540                 config.mps == MLX5_MPW_ENHANCED ? "enhanced " :
2541                 config.mps == MLX5_MPW ? "legacy " : "",
2542                 config.mps != MLX5_MPW_DISABLED ? "enabled" : "disabled");
2543         if (config.cqe_comp && !cqe_comp) {
2544                 DRV_LOG(WARNING, "Rx CQE compression isn't supported");
2545                 config.cqe_comp = 0;
2546         }
2547         if (config.cqe_pad && !cqe_pad) {
2548                 DRV_LOG(WARNING, "Rx CQE padding isn't supported");
2549                 config.cqe_pad = 0;
2550         } else if (config.cqe_pad) {
2551                 DRV_LOG(INFO, "Rx CQE padding is enabled");
2552         }
2553         if (config.devx) {
2554                 priv->counter_fallback = 0;
2555                 err = mlx5_devx_cmd_query_hca_attr(sh->ctx, &config.hca_attr);
2556                 if (err) {
2557                         err = -err;
2558                         goto error;
2559                 }
2560                 if (!config.hca_attr.flow_counters_dump)
2561                         priv->counter_fallback = 1;
2562 #ifndef HAVE_IBV_DEVX_ASYNC
2563                 priv->counter_fallback = 1;
2564 #endif
2565                 if (priv->counter_fallback)
2566                         DRV_LOG(INFO, "Use fall-back DV counter management");
2567                 /* Check for LRO support. */
2568                 if (config.dest_tir && config.hca_attr.lro_cap &&
2569                     config.dv_flow_en) {
2570                         /* TBD check tunnel lro caps. */
2571                         config.lro.supported = config.hca_attr.lro_cap;
2572                         DRV_LOG(DEBUG, "Device supports LRO");
2573                         /*
2574                          * If LRO timeout is not configured by application,
2575                          * use the minimal supported value.
2576                          */
2577                         if (!config.lro.timeout)
2578                                 config.lro.timeout =
2579                                 config.hca_attr.lro_timer_supported_periods[0];
2580                         DRV_LOG(DEBUG, "LRO session timeout set to %d usec",
2581                                 config.lro.timeout);
2582                 }
2583 #if defined(HAVE_MLX5DV_DR) && defined(HAVE_MLX5_DR_CREATE_ACTION_FLOW_METER)
2584                 if (config.hca_attr.qos.sup && config.hca_attr.qos.srtcm_sup &&
2585                     config.dv_flow_en) {
2586                         uint8_t reg_c_mask =
2587                                 config.hca_attr.qos.flow_meter_reg_c_ids;
2588                         /*
2589                          * Meter needs two REG_C's for color match and pre-sfx
2590                          * flow match. Here get the REG_C for color match.
2591                          * REG_C_0 and REG_C_1 is reserved for metadata feature.
2592                          */
2593                         reg_c_mask &= 0xfc;
2594                         if (__builtin_popcount(reg_c_mask) < 1) {
2595                                 priv->mtr_en = 0;
2596                                 DRV_LOG(WARNING, "No available register for"
2597                                         " meter.");
2598                         } else {
2599                                 priv->mtr_color_reg = ffs(reg_c_mask) - 1 +
2600                                                       REG_C_0;
2601                                 priv->mtr_en = 1;
2602                                 priv->mtr_reg_share =
2603                                       config.hca_attr.qos.flow_meter_reg_share;
2604                                 DRV_LOG(DEBUG, "The REG_C meter uses is %d",
2605                                         priv->mtr_color_reg);
2606                         }
2607                 }
2608 #endif
2609         }
2610         if (config.mprq.enabled && mprq) {
2611                 if (config.mprq.stride_num_n > mprq_max_stride_num_n ||
2612                     config.mprq.stride_num_n < mprq_min_stride_num_n) {
2613                         config.mprq.stride_num_n =
2614                                 RTE_MAX(MLX5_MPRQ_STRIDE_NUM_N,
2615                                         mprq_min_stride_num_n);
2616                         DRV_LOG(WARNING,
2617                                 "the number of strides"
2618                                 " for Multi-Packet RQ is out of range,"
2619                                 " setting default value (%u)",
2620                                 1 << config.mprq.stride_num_n);
2621                 }
2622                 config.mprq.min_stride_size_n = mprq_min_stride_size_n;
2623                 config.mprq.max_stride_size_n = mprq_max_stride_size_n;
2624         } else if (config.mprq.enabled && !mprq) {
2625                 DRV_LOG(WARNING, "Multi-Packet RQ isn't supported");
2626                 config.mprq.enabled = 0;
2627         }
2628         if (config.max_dump_files_num == 0)
2629                 config.max_dump_files_num = 128;
2630         eth_dev = rte_eth_dev_allocate(name);
2631         if (eth_dev == NULL) {
2632                 DRV_LOG(ERR, "can not allocate rte ethdev");
2633                 err = ENOMEM;
2634                 goto error;
2635         }
2636         /* Flag to call rte_eth_dev_release_port() in rte_eth_dev_close(). */
2637         eth_dev->data->dev_flags |= RTE_ETH_DEV_CLOSE_REMOVE;
2638         if (priv->representor) {
2639                 eth_dev->data->dev_flags |= RTE_ETH_DEV_REPRESENTOR;
2640                 eth_dev->data->representor_id = priv->representor_id;
2641         }
2642         /*
2643          * Store associated network device interface index. This index
2644          * is permanent throughout the lifetime of device. So, we may store
2645          * the ifindex here and use the cached value further.
2646          */
2647         MLX5_ASSERT(spawn->ifindex);
2648         priv->if_index = spawn->ifindex;
2649         eth_dev->data->dev_private = priv;
2650         priv->dev_data = eth_dev->data;
2651         eth_dev->data->mac_addrs = priv->mac;
2652         eth_dev->device = dpdk_dev;
2653         /* Configure the first MAC address by default. */
2654         if (mlx5_get_mac(eth_dev, &mac.addr_bytes)) {
2655                 DRV_LOG(ERR,
2656                         "port %u cannot get MAC address, is mlx5_en"
2657                         " loaded? (errno: %s)",
2658                         eth_dev->data->port_id, strerror(rte_errno));
2659                 err = ENODEV;
2660                 goto error;
2661         }
2662         DRV_LOG(INFO,
2663                 "port %u MAC address is %02x:%02x:%02x:%02x:%02x:%02x",
2664                 eth_dev->data->port_id,
2665                 mac.addr_bytes[0], mac.addr_bytes[1],
2666                 mac.addr_bytes[2], mac.addr_bytes[3],
2667                 mac.addr_bytes[4], mac.addr_bytes[5]);
2668 #ifdef RTE_LIBRTE_MLX5_DEBUG
2669         {
2670                 char ifname[IF_NAMESIZE];
2671
2672                 if (mlx5_get_ifname(eth_dev, &ifname) == 0)
2673                         DRV_LOG(DEBUG, "port %u ifname is \"%s\"",
2674                                 eth_dev->data->port_id, ifname);
2675                 else
2676                         DRV_LOG(DEBUG, "port %u ifname is unknown",
2677                                 eth_dev->data->port_id);
2678         }
2679 #endif
2680         /* Get actual MTU if possible. */
2681         err = mlx5_get_mtu(eth_dev, &priv->mtu);
2682         if (err) {
2683                 err = rte_errno;
2684                 goto error;
2685         }
2686         DRV_LOG(DEBUG, "port %u MTU is %u", eth_dev->data->port_id,
2687                 priv->mtu);
2688         /* Initialize burst functions to prevent crashes before link-up. */
2689         eth_dev->rx_pkt_burst = removed_rx_burst;
2690         eth_dev->tx_pkt_burst = removed_tx_burst;
2691         eth_dev->dev_ops = &mlx5_dev_ops;
2692         /* Register MAC address. */
2693         claim_zero(mlx5_mac_addr_add(eth_dev, &mac, 0, 0));
2694         if (config.vf && config.vf_nl_en)
2695                 mlx5_nl_mac_addr_sync(priv->nl_socket_route,
2696                                       mlx5_ifindex(eth_dev),
2697                                       eth_dev->data->mac_addrs,
2698                                       MLX5_MAX_MAC_ADDRESSES);
2699         TAILQ_INIT(&priv->flows);
2700         TAILQ_INIT(&priv->ctrl_flows);
2701         TAILQ_INIT(&priv->flow_meters);
2702         TAILQ_INIT(&priv->flow_meter_profiles);
2703         /* Hint libmlx5 to use PMD allocator for data plane resources */
2704         struct mlx5dv_ctx_allocators alctr = {
2705                 .alloc = &mlx5_alloc_verbs_buf,
2706                 .free = &mlx5_free_verbs_buf,
2707                 .data = priv,
2708         };
2709         mlx5_glue->dv_set_context_attr(sh->ctx,
2710                                        MLX5DV_CTX_ATTR_BUF_ALLOCATORS,
2711                                        (void *)((uintptr_t)&alctr));
2712         /* Bring Ethernet device up. */
2713         DRV_LOG(DEBUG, "port %u forcing Ethernet interface up",
2714                 eth_dev->data->port_id);
2715         mlx5_set_link_up(eth_dev);
2716         /*
2717          * Even though the interrupt handler is not installed yet,
2718          * interrupts will still trigger on the async_fd from
2719          * Verbs context returned by ibv_open_device().
2720          */
2721         mlx5_link_update(eth_dev, 0);
2722 #ifdef HAVE_MLX5DV_DR_ESWITCH
2723         if (!(config.hca_attr.eswitch_manager && config.dv_flow_en &&
2724               (switch_info->representor || switch_info->master)))
2725                 config.dv_esw_en = 0;
2726 #else
2727         config.dv_esw_en = 0;
2728 #endif
2729         /* Detect minimal data bytes to inline. */
2730         mlx5_set_min_inline(spawn, &config);
2731         /* Store device configuration on private structure. */
2732         priv->config = config;
2733         /* Create context for virtual machine VLAN workaround. */
2734         priv->vmwa_context = mlx5_vlan_vmwa_init(eth_dev, spawn->ifindex);
2735         if (config.dv_flow_en) {
2736                 err = mlx5_alloc_shared_dr(priv);
2737                 if (err)
2738                         goto error;
2739                 /*
2740                  * RSS id is shared with meter flow id. Meter flow id can only
2741                  * use the 24 MSB of the register.
2742                  */
2743                 priv->qrss_id_pool = mlx5_flow_id_pool_alloc(UINT32_MAX >>
2744                                      MLX5_MTR_COLOR_BITS);
2745                 if (!priv->qrss_id_pool) {
2746                         DRV_LOG(ERR, "can't create flow id pool");
2747                         err = ENOMEM;
2748                         goto error;
2749                 }
2750         }
2751         /* Supported Verbs flow priority number detection. */
2752         err = mlx5_flow_discover_priorities(eth_dev);
2753         if (err < 0) {
2754                 err = -err;
2755                 goto error;
2756         }
2757         priv->config.flow_prio = err;
2758         if (!priv->config.dv_esw_en &&
2759             priv->config.dv_xmeta_en != MLX5_XMETA_MODE_LEGACY) {
2760                 DRV_LOG(WARNING, "metadata mode %u is not supported "
2761                                  "(no E-Switch)", priv->config.dv_xmeta_en);
2762                 priv->config.dv_xmeta_en = MLX5_XMETA_MODE_LEGACY;
2763         }
2764         mlx5_set_metadata_mask(eth_dev);
2765         if (priv->config.dv_xmeta_en != MLX5_XMETA_MODE_LEGACY &&
2766             !priv->sh->dv_regc0_mask) {
2767                 DRV_LOG(ERR, "metadata mode %u is not supported "
2768                              "(no metadata reg_c[0] is available)",
2769                              priv->config.dv_xmeta_en);
2770                         err = ENOTSUP;
2771                         goto error;
2772         }
2773         /*
2774          * Allocate the buffer for flow creating, just once.
2775          * The allocation must be done before any flow creating.
2776          */
2777         mlx5_flow_alloc_intermediate(eth_dev);
2778         /* Query availibility of metadata reg_c's. */
2779         err = mlx5_flow_discover_mreg_c(eth_dev);
2780         if (err < 0) {
2781                 err = -err;
2782                 goto error;
2783         }
2784         if (!mlx5_flow_ext_mreg_supported(eth_dev)) {
2785                 DRV_LOG(DEBUG,
2786                         "port %u extensive metadata register is not supported",
2787                         eth_dev->data->port_id);
2788                 if (priv->config.dv_xmeta_en != MLX5_XMETA_MODE_LEGACY) {
2789                         DRV_LOG(ERR, "metadata mode %u is not supported "
2790                                      "(no metadata registers available)",
2791                                      priv->config.dv_xmeta_en);
2792                         err = ENOTSUP;
2793                         goto error;
2794                 }
2795         }
2796         if (priv->config.dv_flow_en &&
2797             priv->config.dv_xmeta_en != MLX5_XMETA_MODE_LEGACY &&
2798             mlx5_flow_ext_mreg_supported(eth_dev) &&
2799             priv->sh->dv_regc0_mask) {
2800                 priv->mreg_cp_tbl = mlx5_hlist_create(MLX5_FLOW_MREG_HNAME,
2801                                                       MLX5_FLOW_MREG_HTABLE_SZ);
2802                 if (!priv->mreg_cp_tbl) {
2803                         err = ENOMEM;
2804                         goto error;
2805                 }
2806         }
2807         return eth_dev;
2808 error:
2809         if (priv) {
2810                 if (priv->mreg_cp_tbl)
2811                         mlx5_hlist_destroy(priv->mreg_cp_tbl, NULL, NULL);
2812                 if (priv->sh)
2813                         mlx5_free_shared_dr(priv);
2814                 if (priv->nl_socket_route >= 0)
2815                         close(priv->nl_socket_route);
2816                 if (priv->nl_socket_rdma >= 0)
2817                         close(priv->nl_socket_rdma);
2818                 if (priv->vmwa_context)
2819                         mlx5_vlan_vmwa_exit(priv->vmwa_context);
2820                 if (priv->qrss_id_pool)
2821                         mlx5_flow_id_pool_release(priv->qrss_id_pool);
2822                 if (own_domain_id)
2823                         claim_zero(rte_eth_switch_domain_free(priv->domain_id));
2824                 rte_free(priv);
2825                 if (eth_dev != NULL)
2826                         eth_dev->data->dev_private = NULL;
2827         }
2828         if (eth_dev != NULL) {
2829                 /* mac_addrs must not be freed alone because part of dev_private */
2830                 eth_dev->data->mac_addrs = NULL;
2831                 rte_eth_dev_release_port(eth_dev);
2832         }
2833         if (sh)
2834                 mlx5_free_shared_ibctx(sh);
2835         MLX5_ASSERT(err > 0);
2836         rte_errno = err;
2837         return NULL;
2838 }
2839
2840 /**
2841  * Comparison callback to sort device data.
2842  *
2843  * This is meant to be used with qsort().
2844  *
2845  * @param a[in]
2846  *   Pointer to pointer to first data object.
2847  * @param b[in]
2848  *   Pointer to pointer to second data object.
2849  *
2850  * @return
2851  *   0 if both objects are equal, less than 0 if the first argument is less
2852  *   than the second, greater than 0 otherwise.
2853  */
2854 static int
2855 mlx5_dev_spawn_data_cmp(const void *a, const void *b)
2856 {
2857         const struct mlx5_switch_info *si_a =
2858                 &((const struct mlx5_dev_spawn_data *)a)->info;
2859         const struct mlx5_switch_info *si_b =
2860                 &((const struct mlx5_dev_spawn_data *)b)->info;
2861         int ret;
2862
2863         /* Master device first. */
2864         ret = si_b->master - si_a->master;
2865         if (ret)
2866                 return ret;
2867         /* Then representor devices. */
2868         ret = si_b->representor - si_a->representor;
2869         if (ret)
2870                 return ret;
2871         /* Unidentified devices come last in no specific order. */
2872         if (!si_a->representor)
2873                 return 0;
2874         /* Order representors by name. */
2875         return si_a->port_name - si_b->port_name;
2876 }
2877
2878 /**
2879  * Match PCI information for possible slaves of bonding device.
2880  *
2881  * @param[in] ibv_dev
2882  *   Pointer to Infiniband device structure.
2883  * @param[in] pci_dev
2884  *   Pointer to PCI device structure to match PCI address.
2885  * @param[in] nl_rdma
2886  *   Netlink RDMA group socket handle.
2887  *
2888  * @return
2889  *   negative value if no bonding device found, otherwise
2890  *   positive index of slave PF in bonding.
2891  */
2892 static int
2893 mlx5_device_bond_pci_match(const struct ibv_device *ibv_dev,
2894                            const struct rte_pci_device *pci_dev,
2895                            int nl_rdma)
2896 {
2897         char ifname[IF_NAMESIZE + 1];
2898         unsigned int ifindex;
2899         unsigned int np, i;
2900         FILE *file = NULL;
2901         int pf = -1;
2902
2903         /*
2904          * Try to get master device name. If something goes
2905          * wrong suppose the lack of kernel support and no
2906          * bonding devices.
2907          */
2908         if (nl_rdma < 0)
2909                 return -1;
2910         if (!strstr(ibv_dev->name, "bond"))
2911                 return -1;
2912         np = mlx5_nl_portnum(nl_rdma, ibv_dev->name);
2913         if (!np)
2914                 return -1;
2915         /*
2916          * The Master device might not be on the predefined
2917          * port (not on port index 1, it is not garanted),
2918          * we have to scan all Infiniband device port and
2919          * find master.
2920          */
2921         for (i = 1; i <= np; ++i) {
2922                 /* Check whether Infiniband port is populated. */
2923                 ifindex = mlx5_nl_ifindex(nl_rdma, ibv_dev->name, i);
2924                 if (!ifindex)
2925                         continue;
2926                 if (!if_indextoname(ifindex, ifname))
2927                         continue;
2928                 /* Try to read bonding slave names from sysfs. */
2929                 MKSTR(slaves,
2930                       "/sys/class/net/%s/master/bonding/slaves", ifname);
2931                 file = fopen(slaves, "r");
2932                 if (file)
2933                         break;
2934         }
2935         if (!file)
2936                 return -1;
2937         /* Use safe format to check maximal buffer length. */
2938         MLX5_ASSERT(atol(RTE_STR(IF_NAMESIZE)) == IF_NAMESIZE);
2939         while (fscanf(file, "%" RTE_STR(IF_NAMESIZE) "s", ifname) == 1) {
2940                 char tmp_str[IF_NAMESIZE + 32];
2941                 struct rte_pci_addr pci_addr;
2942                 struct mlx5_switch_info info;
2943
2944                 /* Process slave interface names in the loop. */
2945                 snprintf(tmp_str, sizeof(tmp_str),
2946                          "/sys/class/net/%s", ifname);
2947                 if (mlx5_dev_to_pci_addr(tmp_str, &pci_addr)) {
2948                         DRV_LOG(WARNING, "can not get PCI address"
2949                                          " for netdev \"%s\"", ifname);
2950                         continue;
2951                 }
2952                 if (pci_dev->addr.domain != pci_addr.domain ||
2953                     pci_dev->addr.bus != pci_addr.bus ||
2954                     pci_dev->addr.devid != pci_addr.devid ||
2955                     pci_dev->addr.function != pci_addr.function)
2956                         continue;
2957                 /* Slave interface PCI address match found. */
2958                 fclose(file);
2959                 snprintf(tmp_str, sizeof(tmp_str),
2960                          "/sys/class/net/%s/phys_port_name", ifname);
2961                 file = fopen(tmp_str, "rb");
2962                 if (!file)
2963                         break;
2964                 info.name_type = MLX5_PHYS_PORT_NAME_TYPE_NOTSET;
2965                 if (fscanf(file, "%32s", tmp_str) == 1)
2966                         mlx5_translate_port_name(tmp_str, &info);
2967                 if (info.name_type == MLX5_PHYS_PORT_NAME_TYPE_LEGACY ||
2968                     info.name_type == MLX5_PHYS_PORT_NAME_TYPE_UPLINK)
2969                         pf = info.port_name;
2970                 break;
2971         }
2972         if (file)
2973                 fclose(file);
2974         return pf;
2975 }
2976
2977 /**
2978  * DPDK callback to register a PCI device.
2979  *
2980  * This function spawns Ethernet devices out of a given PCI device.
2981  *
2982  * @param[in] pci_drv
2983  *   PCI driver structure (mlx5_driver).
2984  * @param[in] pci_dev
2985  *   PCI device information.
2986  *
2987  * @return
2988  *   0 on success, a negative errno value otherwise and rte_errno is set.
2989  */
2990 static int
2991 mlx5_pci_probe(struct rte_pci_driver *pci_drv __rte_unused,
2992                struct rte_pci_device *pci_dev)
2993 {
2994         struct ibv_device **ibv_list;
2995         /*
2996          * Number of found IB Devices matching with requested PCI BDF.
2997          * nd != 1 means there are multiple IB devices over the same
2998          * PCI device and we have representors and master.
2999          */
3000         unsigned int nd = 0;
3001         /*
3002          * Number of found IB device Ports. nd = 1 and np = 1..n means
3003          * we have the single multiport IB device, and there may be
3004          * representors attached to some of found ports.
3005          */
3006         unsigned int np = 0;
3007         /*
3008          * Number of DPDK ethernet devices to Spawn - either over
3009          * multiple IB devices or multiple ports of single IB device.
3010          * Actually this is the number of iterations to spawn.
3011          */
3012         unsigned int ns = 0;
3013         /*
3014          * Bonding device
3015          *   < 0 - no bonding device (single one)
3016          *  >= 0 - bonding device (value is slave PF index)
3017          */
3018         int bd = -1;
3019         struct mlx5_dev_spawn_data *list = NULL;
3020         struct mlx5_dev_config dev_config;
3021         int ret;
3022
3023         if (mlx5_class_get(pci_dev->device.devargs) != MLX5_CLASS_NET) {
3024                 DRV_LOG(DEBUG, "Skip probing - should be probed by other mlx5"
3025                         " driver.");
3026                 return 1;
3027         }
3028         if (rte_eal_process_type() == RTE_PROC_PRIMARY)
3029                 mlx5_pmd_socket_init();
3030         ret = mlx5_init_once();
3031         if (ret) {
3032                 DRV_LOG(ERR, "unable to init PMD global data: %s",
3033                         strerror(rte_errno));
3034                 return -rte_errno;
3035         }
3036         MLX5_ASSERT(pci_drv == &mlx5_driver);
3037         errno = 0;
3038         ibv_list = mlx5_glue->get_device_list(&ret);
3039         if (!ibv_list) {
3040                 rte_errno = errno ? errno : ENOSYS;
3041                 DRV_LOG(ERR, "cannot list devices, is ib_uverbs loaded?");
3042                 return -rte_errno;
3043         }
3044         /*
3045          * First scan the list of all Infiniband devices to find
3046          * matching ones, gathering into the list.
3047          */
3048         struct ibv_device *ibv_match[ret + 1];
3049         int nl_route = mlx5_nl_init(NETLINK_ROUTE);
3050         int nl_rdma = mlx5_nl_init(NETLINK_RDMA);
3051         unsigned int i;
3052
3053         while (ret-- > 0) {
3054                 struct rte_pci_addr pci_addr;
3055
3056                 DRV_LOG(DEBUG, "checking device \"%s\"", ibv_list[ret]->name);
3057                 bd = mlx5_device_bond_pci_match
3058                                 (ibv_list[ret], pci_dev, nl_rdma);
3059                 if (bd >= 0) {
3060                         /*
3061                          * Bonding device detected. Only one match is allowed,
3062                          * the bonding is supported over multi-port IB device,
3063                          * there should be no matches on representor PCI
3064                          * functions or non VF LAG bonding devices with
3065                          * specified address.
3066                          */
3067                         if (nd) {
3068                                 DRV_LOG(ERR,
3069                                         "multiple PCI match on bonding device"
3070                                         "\"%s\" found", ibv_list[ret]->name);
3071                                 rte_errno = ENOENT;
3072                                 ret = -rte_errno;
3073                                 goto exit;
3074                         }
3075                         DRV_LOG(INFO, "PCI information matches for"
3076                                       " slave %d bonding device \"%s\"",
3077                                       bd, ibv_list[ret]->name);
3078                         ibv_match[nd++] = ibv_list[ret];
3079                         break;
3080                 }
3081                 if (mlx5_dev_to_pci_addr
3082                         (ibv_list[ret]->ibdev_path, &pci_addr))
3083                         continue;
3084                 if (pci_dev->addr.domain != pci_addr.domain ||
3085                     pci_dev->addr.bus != pci_addr.bus ||
3086                     pci_dev->addr.devid != pci_addr.devid ||
3087                     pci_dev->addr.function != pci_addr.function)
3088                         continue;
3089                 DRV_LOG(INFO, "PCI information matches for device \"%s\"",
3090                         ibv_list[ret]->name);
3091                 ibv_match[nd++] = ibv_list[ret];
3092         }
3093         ibv_match[nd] = NULL;
3094         if (!nd) {
3095                 /* No device matches, just complain and bail out. */
3096                 DRV_LOG(WARNING,
3097                         "no Verbs device matches PCI device " PCI_PRI_FMT ","
3098                         " are kernel drivers loaded?",
3099                         pci_dev->addr.domain, pci_dev->addr.bus,
3100                         pci_dev->addr.devid, pci_dev->addr.function);
3101                 rte_errno = ENOENT;
3102                 ret = -rte_errno;
3103                 goto exit;
3104         }
3105         if (nd == 1) {
3106                 /*
3107                  * Found single matching device may have multiple ports.
3108                  * Each port may be representor, we have to check the port
3109                  * number and check the representors existence.
3110                  */
3111                 if (nl_rdma >= 0)
3112                         np = mlx5_nl_portnum(nl_rdma, ibv_match[0]->name);
3113                 if (!np)
3114                         DRV_LOG(WARNING, "can not get IB device \"%s\""
3115                                          " ports number", ibv_match[0]->name);
3116                 if (bd >= 0 && !np) {
3117                         DRV_LOG(ERR, "can not get ports"
3118                                      " for bonding device");
3119                         rte_errno = ENOENT;
3120                         ret = -rte_errno;
3121                         goto exit;
3122                 }
3123         }
3124 #ifndef HAVE_MLX5DV_DR_DEVX_PORT
3125         if (bd >= 0) {
3126                 /*
3127                  * This may happen if there is VF LAG kernel support and
3128                  * application is compiled with older rdma_core library.
3129                  */
3130                 DRV_LOG(ERR,
3131                         "No kernel/verbs support for VF LAG bonding found.");
3132                 rte_errno = ENOTSUP;
3133                 ret = -rte_errno;
3134                 goto exit;
3135         }
3136 #endif
3137         /*
3138          * Now we can determine the maximal
3139          * amount of devices to be spawned.
3140          */
3141         list = rte_zmalloc("device spawn data",
3142                          sizeof(struct mlx5_dev_spawn_data) *
3143                          (np ? np : nd),
3144                          RTE_CACHE_LINE_SIZE);
3145         if (!list) {
3146                 DRV_LOG(ERR, "spawn data array allocation failure");
3147                 rte_errno = ENOMEM;
3148                 ret = -rte_errno;
3149                 goto exit;
3150         }
3151         if (bd >= 0 || np > 1) {
3152                 /*
3153                  * Single IB device with multiple ports found,
3154                  * it may be E-Switch master device and representors.
3155                  * We have to perform identification trough the ports.
3156                  */
3157                 MLX5_ASSERT(nl_rdma >= 0);
3158                 MLX5_ASSERT(ns == 0);
3159                 MLX5_ASSERT(nd == 1);
3160                 MLX5_ASSERT(np);
3161                 for (i = 1; i <= np; ++i) {
3162                         list[ns].max_port = np;
3163                         list[ns].ibv_port = i;
3164                         list[ns].ibv_dev = ibv_match[0];
3165                         list[ns].eth_dev = NULL;
3166                         list[ns].pci_dev = pci_dev;
3167                         list[ns].pf_bond = bd;
3168                         list[ns].ifindex = mlx5_nl_ifindex
3169                                         (nl_rdma, list[ns].ibv_dev->name, i);
3170                         if (!list[ns].ifindex) {
3171                                 /*
3172                                  * No network interface index found for the
3173                                  * specified port, it means there is no
3174                                  * representor on this port. It's OK,
3175                                  * there can be disabled ports, for example
3176                                  * if sriov_numvfs < sriov_totalvfs.
3177                                  */
3178                                 continue;
3179                         }
3180                         ret = -1;
3181                         if (nl_route >= 0)
3182                                 ret = mlx5_nl_switch_info
3183                                                (nl_route,
3184                                                 list[ns].ifindex,
3185                                                 &list[ns].info);
3186                         if (ret || (!list[ns].info.representor &&
3187                                     !list[ns].info.master)) {
3188                                 /*
3189                                  * We failed to recognize representors with
3190                                  * Netlink, let's try to perform the task
3191                                  * with sysfs.
3192                                  */
3193                                 ret =  mlx5_sysfs_switch_info
3194                                                 (list[ns].ifindex,
3195                                                  &list[ns].info);
3196                         }
3197                         if (!ret && bd >= 0) {
3198                                 switch (list[ns].info.name_type) {
3199                                 case MLX5_PHYS_PORT_NAME_TYPE_UPLINK:
3200                                         if (list[ns].info.port_name == bd)
3201                                                 ns++;
3202                                         break;
3203                                 case MLX5_PHYS_PORT_NAME_TYPE_PFVF:
3204                                         if (list[ns].info.pf_num == bd)
3205                                                 ns++;
3206                                         break;
3207                                 default:
3208                                         break;
3209                                 }
3210                                 continue;
3211                         }
3212                         if (!ret && (list[ns].info.representor ^
3213                                      list[ns].info.master))
3214                                 ns++;
3215                 }
3216                 if (!ns) {
3217                         DRV_LOG(ERR,
3218                                 "unable to recognize master/representors"
3219                                 " on the IB device with multiple ports");
3220                         rte_errno = ENOENT;
3221                         ret = -rte_errno;
3222                         goto exit;
3223                 }
3224         } else {
3225                 /*
3226                  * The existence of several matching entries (nd > 1) means
3227                  * port representors have been instantiated. No existing Verbs
3228                  * call nor sysfs entries can tell them apart, this can only
3229                  * be done through Netlink calls assuming kernel drivers are
3230                  * recent enough to support them.
3231                  *
3232                  * In the event of identification failure through Netlink,
3233                  * try again through sysfs, then:
3234                  *
3235                  * 1. A single IB device matches (nd == 1) with single
3236                  *    port (np=0/1) and is not a representor, assume
3237                  *    no switch support.
3238                  *
3239                  * 2. Otherwise no safe assumptions can be made;
3240                  *    complain louder and bail out.
3241                  */
3242                 np = 1;
3243                 for (i = 0; i != nd; ++i) {
3244                         memset(&list[ns].info, 0, sizeof(list[ns].info));
3245                         list[ns].max_port = 1;
3246                         list[ns].ibv_port = 1;
3247                         list[ns].ibv_dev = ibv_match[i];
3248                         list[ns].eth_dev = NULL;
3249                         list[ns].pci_dev = pci_dev;
3250                         list[ns].pf_bond = -1;
3251                         list[ns].ifindex = 0;
3252                         if (nl_rdma >= 0)
3253                                 list[ns].ifindex = mlx5_nl_ifindex
3254                                         (nl_rdma, list[ns].ibv_dev->name, 1);
3255                         if (!list[ns].ifindex) {
3256                                 char ifname[IF_NAMESIZE];
3257
3258                                 /*
3259                                  * Netlink failed, it may happen with old
3260                                  * ib_core kernel driver (before 4.16).
3261                                  * We can assume there is old driver because
3262                                  * here we are processing single ports IB
3263                                  * devices. Let's try sysfs to retrieve
3264                                  * the ifindex. The method works for
3265                                  * master device only.
3266                                  */
3267                                 if (nd > 1) {
3268                                         /*
3269                                          * Multiple devices found, assume
3270                                          * representors, can not distinguish
3271                                          * master/representor and retrieve
3272                                          * ifindex via sysfs.
3273                                          */
3274                                         continue;
3275                                 }
3276                                 ret = mlx5_get_master_ifname
3277                                         (ibv_match[i]->ibdev_path, &ifname);
3278                                 if (!ret)
3279                                         list[ns].ifindex =
3280                                                 if_nametoindex(ifname);
3281                                 if (!list[ns].ifindex) {
3282                                         /*
3283                                          * No network interface index found
3284                                          * for the specified device, it means
3285                                          * there it is neither representor
3286                                          * nor master.
3287                                          */
3288                                         continue;
3289                                 }
3290                         }
3291                         ret = -1;
3292                         if (nl_route >= 0)
3293                                 ret = mlx5_nl_switch_info
3294                                                (nl_route,
3295                                                 list[ns].ifindex,
3296                                                 &list[ns].info);
3297                         if (ret || (!list[ns].info.representor &&
3298                                     !list[ns].info.master)) {
3299                                 /*
3300                                  * We failed to recognize representors with
3301                                  * Netlink, let's try to perform the task
3302                                  * with sysfs.
3303                                  */
3304                                 ret =  mlx5_sysfs_switch_info
3305                                                 (list[ns].ifindex,
3306                                                  &list[ns].info);
3307                         }
3308                         if (!ret && (list[ns].info.representor ^
3309                                      list[ns].info.master)) {
3310                                 ns++;
3311                         } else if ((nd == 1) &&
3312                                    !list[ns].info.representor &&
3313                                    !list[ns].info.master) {
3314                                 /*
3315                                  * Single IB device with
3316                                  * one physical port and
3317                                  * attached network device.
3318                                  * May be SRIOV is not enabled
3319                                  * or there is no representors.
3320                                  */
3321                                 DRV_LOG(INFO, "no E-Switch support detected");
3322                                 ns++;
3323                                 break;
3324                         }
3325                 }
3326                 if (!ns) {
3327                         DRV_LOG(ERR,
3328                                 "unable to recognize master/representors"
3329                                 " on the multiple IB devices");
3330                         rte_errno = ENOENT;
3331                         ret = -rte_errno;
3332                         goto exit;
3333                 }
3334         }
3335         MLX5_ASSERT(ns);
3336         /*
3337          * Sort list to probe devices in natural order for users convenience
3338          * (i.e. master first, then representors from lowest to highest ID).
3339          */
3340         qsort(list, ns, sizeof(*list), mlx5_dev_spawn_data_cmp);
3341         /* Default configuration. */
3342         dev_config = (struct mlx5_dev_config){
3343                 .hw_padding = 0,
3344                 .mps = MLX5_ARG_UNSET,
3345                 .dbnc = MLX5_ARG_UNSET,
3346                 .rx_vec_en = 1,
3347                 .txq_inline_max = MLX5_ARG_UNSET,
3348                 .txq_inline_min = MLX5_ARG_UNSET,
3349                 .txq_inline_mpw = MLX5_ARG_UNSET,
3350                 .txqs_inline = MLX5_ARG_UNSET,
3351                 .vf_nl_en = 1,
3352                 .mr_ext_memseg_en = 1,
3353                 .mprq = {
3354                         .enabled = 0, /* Disabled by default. */
3355                         .stride_num_n = MLX5_MPRQ_STRIDE_NUM_N,
3356                         .max_memcpy_len = MLX5_MPRQ_MEMCPY_DEFAULT_LEN,
3357                         .min_rxqs_num = MLX5_MPRQ_MIN_RXQS,
3358                 },
3359                 .dv_esw_en = 1,
3360                 .dv_flow_en = 1,
3361         };
3362         /* Device specific configuration. */
3363         switch (pci_dev->id.device_id) {
3364         case PCI_DEVICE_ID_MELLANOX_CONNECTX4VF:
3365         case PCI_DEVICE_ID_MELLANOX_CONNECTX4LXVF:
3366         case PCI_DEVICE_ID_MELLANOX_CONNECTX5VF:
3367         case PCI_DEVICE_ID_MELLANOX_CONNECTX5EXVF:
3368         case PCI_DEVICE_ID_MELLANOX_CONNECTX5BFVF:
3369         case PCI_DEVICE_ID_MELLANOX_CONNECTX6VF:
3370         case PCI_DEVICE_ID_MELLANOX_CONNECTX6DXVF:
3371                 dev_config.vf = 1;
3372                 break;
3373         default:
3374                 break;
3375         }
3376         for (i = 0; i != ns; ++i) {
3377                 uint32_t restore;
3378
3379                 list[i].eth_dev = mlx5_dev_spawn(&pci_dev->device,
3380                                                  &list[i],
3381                                                  dev_config);
3382                 if (!list[i].eth_dev) {
3383                         if (rte_errno != EBUSY && rte_errno != EEXIST)
3384                                 break;
3385                         /* Device is disabled or already spawned. Ignore it. */
3386                         continue;
3387                 }
3388                 restore = list[i].eth_dev->data->dev_flags;
3389                 rte_eth_copy_pci_info(list[i].eth_dev, pci_dev);
3390                 /* Restore non-PCI flags cleared by the above call. */
3391                 list[i].eth_dev->data->dev_flags |= restore;
3392                 mlx5_dev_interrupt_handler_devx_install(list[i].eth_dev);
3393                 rte_eth_dev_probing_finish(list[i].eth_dev);
3394         }
3395         if (i != ns) {
3396                 DRV_LOG(ERR,
3397                         "probe of PCI device " PCI_PRI_FMT " aborted after"
3398                         " encountering an error: %s",
3399                         pci_dev->addr.domain, pci_dev->addr.bus,
3400                         pci_dev->addr.devid, pci_dev->addr.function,
3401                         strerror(rte_errno));
3402                 ret = -rte_errno;
3403                 /* Roll back. */
3404                 while (i--) {
3405                         if (!list[i].eth_dev)
3406                                 continue;
3407                         mlx5_dev_close(list[i].eth_dev);
3408                         /* mac_addrs must not be freed because in dev_private */
3409                         list[i].eth_dev->data->mac_addrs = NULL;
3410                         claim_zero(rte_eth_dev_release_port(list[i].eth_dev));
3411                 }
3412                 /* Restore original error. */
3413                 rte_errno = -ret;
3414         } else {
3415                 ret = 0;
3416         }
3417 exit:
3418         /*
3419          * Do the routine cleanup:
3420          * - close opened Netlink sockets
3421          * - free allocated spawn data array
3422          * - free the Infiniband device list
3423          */
3424         if (nl_rdma >= 0)
3425                 close(nl_rdma);
3426         if (nl_route >= 0)
3427                 close(nl_route);
3428         if (list)
3429                 rte_free(list);
3430         MLX5_ASSERT(ibv_list);
3431         mlx5_glue->free_device_list(ibv_list);
3432         return ret;
3433 }
3434
3435 /**
3436  * Look for the ethernet device belonging to mlx5 driver.
3437  *
3438  * @param[in] port_id
3439  *   port_id to start looking for device.
3440  * @param[in] pci_dev
3441  *   Pointer to the hint PCI device. When device is being probed
3442  *   the its siblings (master and preceding representors might
3443  *   not have assigned driver yet (because the mlx5_pci_probe()
3444  *   is not completed yet, for this case match on hint PCI
3445  *   device may be used to detect sibling device.
3446  *
3447  * @return
3448  *   port_id of found device, RTE_MAX_ETHPORT if not found.
3449  */
3450 uint16_t
3451 mlx5_eth_find_next(uint16_t port_id, struct rte_pci_device *pci_dev)
3452 {
3453         while (port_id < RTE_MAX_ETHPORTS) {
3454                 struct rte_eth_dev *dev = &rte_eth_devices[port_id];
3455
3456                 if (dev->state != RTE_ETH_DEV_UNUSED &&
3457                     dev->device &&
3458                     (dev->device == &pci_dev->device ||
3459                      (dev->device->driver &&
3460                      dev->device->driver->name &&
3461                      !strcmp(dev->device->driver->name, MLX5_DRIVER_NAME))))
3462                         break;
3463                 port_id++;
3464         }
3465         if (port_id >= RTE_MAX_ETHPORTS)
3466                 return RTE_MAX_ETHPORTS;
3467         return port_id;
3468 }
3469
3470 /**
3471  * DPDK callback to remove a PCI device.
3472  *
3473  * This function removes all Ethernet devices belong to a given PCI device.
3474  *
3475  * @param[in] pci_dev
3476  *   Pointer to the PCI device.
3477  *
3478  * @return
3479  *   0 on success, the function cannot fail.
3480  */
3481 static int
3482 mlx5_pci_remove(struct rte_pci_device *pci_dev)
3483 {
3484         uint16_t port_id;
3485
3486         RTE_ETH_FOREACH_DEV_OF(port_id, &pci_dev->device)
3487                 rte_eth_dev_close(port_id);
3488         return 0;
3489 }
3490
3491 static const struct rte_pci_id mlx5_pci_id_map[] = {
3492         {
3493                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
3494                                PCI_DEVICE_ID_MELLANOX_CONNECTX4)
3495         },
3496         {
3497                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
3498                                PCI_DEVICE_ID_MELLANOX_CONNECTX4VF)
3499         },
3500         {
3501                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
3502                                PCI_DEVICE_ID_MELLANOX_CONNECTX4LX)
3503         },
3504         {
3505                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
3506                                PCI_DEVICE_ID_MELLANOX_CONNECTX4LXVF)
3507         },
3508         {
3509                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
3510                                PCI_DEVICE_ID_MELLANOX_CONNECTX5)
3511         },
3512         {
3513                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
3514                                PCI_DEVICE_ID_MELLANOX_CONNECTX5VF)
3515         },
3516         {
3517                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
3518                                PCI_DEVICE_ID_MELLANOX_CONNECTX5EX)
3519         },
3520         {
3521                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
3522                                PCI_DEVICE_ID_MELLANOX_CONNECTX5EXVF)
3523         },
3524         {
3525                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
3526                                PCI_DEVICE_ID_MELLANOX_CONNECTX5BF)
3527         },
3528         {
3529                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
3530                                PCI_DEVICE_ID_MELLANOX_CONNECTX5BFVF)
3531         },
3532         {
3533                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
3534                                 PCI_DEVICE_ID_MELLANOX_CONNECTX6)
3535         },
3536         {
3537                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
3538                                 PCI_DEVICE_ID_MELLANOX_CONNECTX6VF)
3539         },
3540         {
3541                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
3542                                 PCI_DEVICE_ID_MELLANOX_CONNECTX6DX)
3543         },
3544         {
3545                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
3546                                 PCI_DEVICE_ID_MELLANOX_CONNECTX6DXVF)
3547         },
3548         {
3549                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
3550                                 PCI_DEVICE_ID_MELLANOX_CONNECTX6DXBF)
3551         },
3552         {
3553                 .vendor_id = 0
3554         }
3555 };
3556
3557 static struct rte_pci_driver mlx5_driver = {
3558         .driver = {
3559                 .name = MLX5_DRIVER_NAME
3560         },
3561         .id_table = mlx5_pci_id_map,
3562         .probe = mlx5_pci_probe,
3563         .remove = mlx5_pci_remove,
3564         .dma_map = mlx5_dma_map,
3565         .dma_unmap = mlx5_dma_unmap,
3566         .drv_flags = RTE_PCI_DRV_INTR_LSC | RTE_PCI_DRV_INTR_RMV |
3567                      RTE_PCI_DRV_PROBE_AGAIN,
3568 };
3569
3570 /**
3571  * Driver initialization routine.
3572  */
3573 RTE_INIT(rte_mlx5_pmd_init)
3574 {
3575         /* Initialize driver log type. */
3576         mlx5_logtype = rte_log_register("pmd.net.mlx5");
3577         if (mlx5_logtype >= 0)
3578                 rte_log_set_level(mlx5_logtype, RTE_LOG_NOTICE);
3579
3580         /* Build the static tables for Verbs conversion. */
3581         mlx5_set_ptype_table();
3582         mlx5_set_cksum_table();
3583         mlx5_set_swp_types_table();
3584         if (mlx5_glue)
3585                 rte_pci_register(&mlx5_driver);
3586 }
3587
3588 RTE_PMD_EXPORT_NAME(net_mlx5, __COUNTER__);
3589 RTE_PMD_REGISTER_PCI_TABLE(net_mlx5, mlx5_pci_id_map);
3590 RTE_PMD_REGISTER_KMOD_DEP(net_mlx5, "* ib_uverbs & mlx5_core & mlx5_ib");